How each number here is made

In one read we look at a shelf a few times. The share of those looks that found a product on sale is what we publish, and we call it in stock at read. Two looks in a day cannot carry a daily percentage: on a single day the margin of error is close to 49 points either way, and nobody should act on that. A 7-day figure needs at least 4 reads across a week, and a 30-day figure at least 20. Until a rate has that behind it, the screen says "not enough reads yet" and prints the counts it has.

Each look falls in a morning window (05:00 to 13:59 IST) or an evening one (14:00 to 04:59). When a read holds more evening windows than morning ones, a morning against evening split is two counts and no pattern, and the screen says so.

We do not print a percentage without the count behind it. Every figure carries the stores we read out of the stores on our list, the shelves we read out of the shelves we set out to read, how many of those were read to the end, and how many attempts failed and why. The line under a figure prints these. An exported file also carries a reference for the exact set of data it came from. Quote it when you ask for a correction.

Sometimes an app does not answer: it asks for a check, refuses, sends an empty page or takes too long. We record that as a failed attempt, with its reason. A failed attempt is left out of every count, and is never used as a sign that a product is missing.

A shelf counts as "read to the end" only where the read carries its own proof. On Blinkit and BigBasket a list states how many products it holds, and the shelf counts when we collected that many. A Blinkit list that states no number counts only when it said there was no more and held fewer than 170 products, because the app cuts a list near 200 however long the category is. A BigBasket list that states no number does not count. On Zepto no shelf counts (see below). We tested this rule against the data: where a shelf really was read to the end, 3 to 4% of its listings were not seen again two weeks later; where it only looked that way, 9 to 27% were not, because the "lost" products had never been looked at. Only a shelf read to the end lets us say "if the product were listed here, we would have seen it".

We say a brand is missing from a store only where a shelf of its category was read to the end in that store. Where a list was cut short, a second page was refused or the app showed only part of the shelf, we read the store and cannot speak for it: it is drawn as "not proven either way". This never applies on Zepto, Swiggy Instamart or Flipkart Minutes, where no shelf can be read to the end.

We read Zepto two ways, and neither shows a whole shelf. A category page shows part of one: on pages that ended before the limit, asking stores directly found that about three in five of their listings were ones the pages had not shown. Asking a store directly works only for products whose id we can work out; where a store lists the same product under one of the app’s random ids, it answers with nothing. So no store on Zepto is counted as one where you are missing, and no listing as lost. What we did read there (where a product is listed, its price and its stock) is shown as read, and every count of it is "at least".

Zepto’s category pages show at most 32 products of a sub-category, so we also ask each store directly about every product we can reach that way, and the store answers with every one of them it lists, in stock or not. About half of the listings stores returned were in stock and half were not; most of the second kind never appear on a category page. Where a page and a direct answer covered the same product in the same store, they agreed on its stock 93 times in 100. So Zepto shows far more listings, and a lower share in stock, than its category pages alone would. Both are what the app answered. We do not compare either figure with another app.

Each read is compared with the previous one. A product that a store listed at the previous read, and that this read did not see, is "looked for, not found" only where its shelf was read to the end in that store this time, with products on it. Anything less is a listing we did not see, and it is not counted anywhere. One read is not enough to call a product delisted: it can come back, or be on another shelf we did not reach. It is never said on Zepto, Swiggy Instamart or Flipkart Minutes.

The apps show a category a page at a time, and some stop early: a Blinkit category list stops near 200 products, a Zepto page at 32, and a BigBasket list can end before the number of products it states. A list that was cut is not a shelf read to the end, and a product missing from it proves nothing. Only a shelf read to the end can show that a product is missing, and calling a product delisted needs every shelf it was on read to the end in two windows a day apart.

A store counts as a gap for your brand only when at least one shelf of a category you sell in was read to the end in that store, and your brand was not on it. Every gap figure is counted over those stores, and not over every store we read. Rivals are then ranked by how many of those stores they are in. Zepto, Swiggy Instamart and Flipkart Minutes never produce a gap: none of them shows a whole shelf, so on those apps "seen in N stores" is a minimum and nothing is said about where a brand is missing. The pincode and the store list come from the same data on the plans that include them, and the count is the same on every plan.

Your categories are the ones holding at least 3% of your listings on that app, and they decide which shelves count for you. Within them, every store read to the end is one of four: both brands on the shelf, you and not the rival, the rival and not you (the gap), or neither. The four add up to the stores read to the end.

A brand that is in nearly every store of an app has almost no store where a rival is on the shelf and it is not. What it does have is gaps in its own range: a product listed in some of the stores that list the brand, and not in others. For each product we take its base: the stores that list any product of the brand, and where that product’s shelf was read to the end in this read. Then we count how many of those stores list the product and how many do not. We look for a product on the shelf most of its stores list it under. A shelf the app says a store does not carry is left out: that is the store’s range, and not a missing product.

The count of stores missing is exact for the stores in the base and a minimum for the app as a whole, because a store whose shelf was cut short is not counted either way. A product listed in at least half of its base is a wide listing, and a store without it is one to raise. A product in fewer than half is a narrow listing (a regional pack, a launch, a test): most of its base leaves it out by design, so it is shown apart. On Zepto, Swiggy Instamart and Flipkart Minutes no shelf is read to the end, so no product is counted as missing there.

A rate with fewer than 20 looks over 5 stores behind it is held back, and the screen prints the counts it has, because a percentage built on a dozen looks is a coin toss. A 7-day rate needs 4 reads within 7 days, and a 30-day rate needs 20. These fill in as the reads arrive.

Besides the grocery apps we read the apps that sell fast from their own dark stores: Myntra M-Now, Nykaa Now, Zepto Pharmacy, Plazza, Supertails, OZi, Slikk, ZILO, KNOT, Peeko, FirstClub, Zepto Cafe, Bistro and Swish, and DMart Ready, DMart’s online shop, which delivers next day or in slots and so is not quick commerce. Each store is read on its own, from a place it delivers to, and its search is read at every store, as a guest except on FirstClub. They are counted on their own: nothing read on them is added to the grocery apps’ figures. Counts are a minimum, as everywhere. Each app states stock its own way, and we print it the way it is stated: units left, in stock or not, or, on Supertails, whether the store would send the product now. A paid figure appears only for an app that marks a paid result; an app that marks none has a blank, never a zero. Cafe apps sell their own food, so their dishes are counted by dish and section, never as a brand. Categories on these apps are ours: placed from the app’s own category words, and from the product’s name where those say nothing, on the same list of categories as the other apps.

FirstClub answers only a signed-in account, so it is read with one; nothing is ordered or kept on it, and each store is read by its own location record. Each store is read page by page in the app’s own order, and its search finds what the category pages do not show. DMart Ready is read as a guest: each online store is the one DMart names for a pincode, its category pages are read in their own order, and its search at every store; a product seen at other stores that a store’s lists did not show is then asked of that store by its id, and such a row has no list position. DMart marks its sponsored cards, and they count as paid. A product DMart lists as coming soon is not on sale at that store yet: it is counted apart and left out of the store’s range. Peeko shows a guest one store and does not say which, so its figures are one Bengaluru store, not tied to an address. KNOT reorders its lists on every visit, so a position there is one visit’s order. Zepto serves each address from two stores, a grocery dark store and a larger store that also sells fashion, home, kitchen, toys, beauty and electronics. The larger stores are read with the category apps, through the app’s store lookup: each store is asked for every product by its id, so a product counts there whether it is in stock or not, and a lookup has no list position and no paid mark. They are kept apart from Zepto’s grocery stores and never added to them.

What your plan does not include is removed on our server before anything reaches your browser: cities outside your plan, brands you do not watch, and the store lists and pincodes of the plans above yours. A count is never changed by this. A locked row on a screen states the real number behind it.

Swiggy Instamart shows a ranked part of a category in place of the whole list, and changes which part it shows. So its counts of stores are a minimum ("seen in at least N stores"), and it never produces a gap or a "not found". A position on Instamart is the place a listing had in the part we were shown, and is not a search rank. A listing that moved 5 or more places inside the first 30 between two looks at the same shelf is recorded as "position moved".

We know each Swiggy Instamart store by its area, and none carries a pincode of its own. A pincode shown beside one tells you where the store is. It proves nothing about what that store lacks, because we only ever see part of its shelf.

We never reach the end of a list on Flipkart Minutes. The first read took one page per store; later reads went further down each list without reaching its end. So it tells us where a product is listed and at what price, and never that a product is missing.

Zepto states a quantity only for products that are in stock, so counting zero quantities would always give 0.0% out of stock. We use the app’s own in-stock mark for availability, and treat the quantity as a sign of depth where it is given. Zepto also decides which store serves a location, so we file a shelf under the store that served it, and not under the location we asked from.

On Zepto a second store sometimes serves the same location as the store we asked for. Products served by the second store are filed under it, and a note says so wherever they appear, so the same shelf is never counted twice.

A read runs over more than one day, so two prices for one product may be two stores, or one store on two days. Unless every look fell inside one morning or evening window, we show the spread as the most that place alone could explain, and print the number of hours the looks cover. The median price is the middle of every price we read for the product across stores in the read. "Prices seen" is how many different prices were displayed. "At the most common price" is the share of stores showing the single most common one.

The same rule covers "stores moved together": only a count taken inside one window says that stores changed at the same time.

Your median product price divided by the category’s median on the same app, times 100. It is a position and not a discount: 100 is the category’s median, and 90 is a tenth below it. "Price vs rival" is the same sum between your median in the category and the rival’s. Both are medians over a category and not matched packs, and neither is compared across apps, because each app sets prices for its own shelf.

Every discount figure here compares the selling price with the MRP the app itself displays, which is a list price the app controls. "Below MRP" is the median, across your products on one app and category, of MRP minus price as a share of MRP. "On promo" is the share of your listings that show an MRP and are priced 10% or more below it. Flipkart Minutes shows a discount on nearly every product, which says more about how it presents prices than about the prices.

We read across every category before going deep into any one, so a count of products is what that read reached. It is not the size of an app’s catalogue, or of yours.

BigBasket publishes a barcode on about 30% of its products, and that is the only exact key any app gives us. A pair matched on a valid barcode is exact. Without one we match on the brand, the pack size and the words of the name, and keep how sure we are with the pair: strong, probable, name only, or commodity. A match we worked out is never shown as exact. Where one listing is the best match for another and the reverse is not true, the pair is shown as a conflict and not quietly settled. "Same pack, different app" compares the median store price of one matched pack on each app, and only matched packs count.

A listing is one product on the shelf of one store: a brand with 10 products in 100 stores has up to 1,000 listings. It is not a physical facing on a rack. Share of shelf is your listings in one category on one app, divided by all the listings in that category on that app, over the stores we read. It is not added across apps, and it is not market share or sales.

Store positions come from the apps themselves. On the brand screens a store is drawn as the 1 km square it is in, and no latitude or longitude, store name or address is shown there: they are shown only in Location intelligence, to an account that holds it (the next rule). Flipkart Minutes, Amazon Now and JioMart are counted by city on the brand screens, and their stores are not listed there. Every Swiggy Instamart store is known by its area and has no pincode.

Location intelligence lists every store of the 22 apps we read, with every detail of it that is public: where it is and how exactly, its address, its area and pincode, the hours and the delivery promise the app states, what the app uses it for, its print station and its range. We find the stores by a census of each app: every week for the grocery apps, and once a month for now for the category apps. A position is exact (within 250 m), approximate (within 1 km), or the area only (further, or never measured), and a street address is given only with an exact position. A store the app lists and nothing placed is listed with no position. Totals that include JioMart and Amazon Now count stores, not only dark stores.

Each refresh is set against the one before it. A store is new in the census the first time we see it: it may be older than that, and it is called likely newly opened only where the app’s own store ids say so. A store is called closed only when two readings on different days agree; after one reading without it, it is not answering, waiting for the second. A move is a position that changed by more than 250 m, said with its distance: a move, or a sharper reading of the same store. These details leave Shelf only on the Location screens and in its files, and only to an account that holds Location intelligence.

At a first read there is nothing to compare with, so every product is "first seen" and not "new", and nothing is called missing. Changes inside one read are real: we look at some shelves twice, and about 7% of the listings we looked at twice had changed in between. A change of that kind is marked "same read": caught between two looks inside one read, and not from one read to the next.

A product we did not see in this read is only that: not seen. Unless every shelf it was on was read to the end in two windows a day apart, and it was on none of them, we say nothing about whether it is still listed.

When a read covers more stores or more categories than the previous one, some of what looks new is new to us and not newly listed. A product is "first seen" the first time we read it. It is called a "new product" only when it is on a shelf read to the end this time, and was not on the same shelf, read to the end, the previous time.

Restock lag is the hours between a listing going out of stock and coming back, measured only for stockouts that came back inside the read. A stockout that had not come back by our last look at that store is "still out at last read": it may have come back after we left, so it is counted apart and kept out of the median. The median is held back below 20 completed stockouts over 5 stores.

In the shelf read, Swiggy Instamart is the one app whose category pages carried a "sponsored" label in what we read. Paid slots in search results are covered under Ads below. The sponsored share is the looks that carried that label divided by all looks, per category and city. It measures how often the label was shown. It says nothing about money, so no rupee figure is shown, and it is not share of voice. "The first page" is the first ten positions of the part of the category Instamart showed us at the last read.

One search is one keyword typed at one store of one city, as a guest (not signed in), with every result kept in the order it appeared: its position, whether the app marked it as paid, the product, its price and whether it was in stock. A city is one to three of its stores (the first near the middle of the city, the others spread apart), and not every store in the city. The keywords are plain words such as "curd", brand names, a brand with a product, and needs such as "evening snacks", one list for every city. This is what a guest sees, and a signed-in shopper may see a different order. We have run the searches once so far, so nothing on the Search or Ads screens is a trend, and most of it was read between 9 pm and 8 am. A store that was closed, and a search that returned paid products and nothing else, are counted and left out of every figure. "Page as read" is one of those searches shown as it came back, with its time. "Out of stock in search" is the list the app itself printed under a search, of products that store could not supply.

Each figure states the searches and the cities behind it. A results page changes from one search to the next, so the figure to rely on is a share over many searches, and not a single page.

The first 10 is the first ten unpaid results of a search. A brand’s share of the first 10 for a keyword is the part of all those unpaid places, across every search of that keyword on one app, that its products held. It is not added across apps. Best unpaid rank is the highest unpaid position any of the brand’s products held in a search, and across many searches we print the middle value. A paid slot has no unpaid rank.

Two things on the Search screen are our own analysis, and are labelled as ours. A keyword is relevant to a brand when the brand holds part of its first 10, or sells in the category that fills that keyword’s first 10 on that app. Where an app states no category, a brand that keeps coming up just below the first 10, on a kind of keyword it already ranks for, counts too. A keyword gap is a relevant keyword where the brand held none of the first 10 in any search we read. "Brands beside you" are the brands that most often share a first 10 with yours on plain keywords, leaving out sister brands of the same company.

Share of search is counted for one kind of keyword (dairy, snacks, personal care: the kinds our keyword list uses) on one app. Over the plain keywords of that kind we add up every unpaid place in the first 10 of every search, and a brand’s share is the part of those places its products held. Brand-name keywords are left out, or a brand would lead its kind by owning its own name. Putting brands in order of that share gives a rank, printed beside the number of brands ranked; only brands that held at least three such places in the kind are counted. "In the first 3" is the share of searches where one of the brand’s products was among the first three unpaid results.

It is a share of places in the results over many searches, read as a guest, mostly at night, in one round of searches. It is not a share of shoppers, clicks or sales. It is not added across apps, and it will change when we search again.

The "What to do" list on the Search screen picks out keywords where the searches we read point to something a brand could act on. It measures nothing new: every line is made of counts shown elsewhere on the Search and Ads screens. There are eight kinds. Defend your name: another brand held a paid slot on searches for your name in five or more cities. Not in the first 10: a keyword relevant to you where none of the first 10 results is yours. Almost there: you came up in at least three in ten searches of a keyword, were in the first 10 in fewer than half, and your best result was usually at rank 11 to 20. City holes: on a keyword where you hold a twentieth or more of the first 10 and appear in at least half the cities read, five or more cities showed none of your products. Another brand’s paid slot where you are in the first 3: on a plain keyword you are usually among the first 3 unpaid results and held no paid slot in five or more cities, and another brand held one in five or more. It says who held a paid slot on the keyword, and not that the slot was above your result. Paid where already in the first 3: you held a paid slot in five or more cities on a plain keyword where your products are already in the first 3 in at least half the searches. A paid slot only: you held a paid slot in five or more cities on a plain keyword where none of the first 10 unpaid results is yours; your products may still come up lower on the page. Unavailable in search: the app listed your product under a search and marked it unavailable at that store.

The order is a fixed sum, the same for every brand: a weight for the kind (1.0 defend your name, 0.9 not in the first 10, 0.8 almost there, 0.7 another brand’s paid slot where you are in the first 3, 0.6 city holes, 0.5 paid where already in the first 3, 0.5 unavailable in search, 0.45 a paid slot only), times the keyword’s place in our list (a head keyword counts in full, a long-tail one half), times a figure between a half and one for how much of the keyword we read. On the screen no kind takes more than two of the first six lines. The list states what was on the page. It does not say what a rival intended, and it does not suggest a bid or a budget, because we see where slots appeared and not ad accounts.

Search shows where a brand’s products rank unpaid. Ads shows where a brand held a paid slot. Put together, every plain keyword a brand came up on falls into one of a few groups, and we count them. Unpaid rank has three bands, each with one rule. Usually in the first 3: the brand’s best unpaid result was in the first three in at least half the searches of that keyword. Lower on the page: in the first 10 in some searches, and in the first three in fewer than half. None of the first 10: no unpaid place in the first 10 in any search we read. A paid slot counts once it was seen in five of the cities read, the same rule as every other table of paid slots; one seen in fewer cities is counted and not treated as a finding. Brand-name keywords are left out: a paid slot on a brand’s own name is a defence, and other brands’ names have their own view.

Three lists come out of it: keywords where a brand held a paid slot while already in the first 3; keywords where it held a paid slot and none of the first 10 unpaid results is its own; and keywords where a brand is usually in the first 3 with no paid slot of its own in five or more cities, while another brand held one there (named from five cities, and never a sister brand). The last one says who held a paid slot on the keyword and in how many cities. It does not say the slot was above the brand’s own result, because the two positions are not compared search by search. The same is worked out for a rival you watch and set beside yours, keyword by keyword. This is our own analysis of what was on the page, over many searches, read as a guest. It states where results and paid slots appeared. It does not say what a brand chose, what a slot cost or what it earned, and a paid slot can serve a purpose we cannot see from the page. How often a page opens with a paid slot depends on the app’s layout and differs from app to app, so it is printed as a fact about a keyword and not about a brand.

For every keyword we read on an app, the finder sets how crowded its first 10 is beside the signs of demand the app shows. How crowded is read from the unpaid first 10 across the searches read: the leading brand’s share, the share of the first three brands, and, where the app’s paid slots were read, how many brands held a paid slot in five or more cities. A keyword is open when it is in the least crowded third of that app’s keywords, crowded in the most crowded third, and contested between. It compares a keyword with the others on the same app; it is not a measure of cost, and it is never added across apps.

No app publishes how often a word is searched, so we do not print a search volume. Demand is read from three things the apps show: how many brands held a paid slot on the keyword in five or more cities (the apps place paid slots on the searches shoppers make), whether our keyword list counts it a head keyword, a body keyword or a long-tail one (a judgement made when the list was written), and how often the app itself listed products that shoppers searched for and could not get. Strong, moderate and light are the keyword’s third of the app’s keywords by those signs together. Each keyword also names its first five brands by share of the first 10, with the middle value of each one’s best unpaid rank, so where every brand stands on the keyword can be read in one row. “Worth a try” is a filter of ours, the same for every brand: plain keywords in the brand’s main categories (those holding at least a tenth of the plain keywords it has a first-10 slot on), open or contested, with moderate or strong signs of demand, where the brand is not usually in the first 3 (its best unpaid result in the first three in under half the searches read).

Product research sets every category of an app beside the others, for a brand choosing what to launch next and for anyone studying a market before entering it. It measures nothing new: every figure is one the market, demand and store steps already counted, from the same read. How crowded a shelf is comes from five things, each a category’s place among the others on the same app: the share of listings held by the three largest brands, the brands for each place on a typical store’s shelf (the category’s brands over the products a typical store lists in it), the share of listings on the app’s own labels, the median discount off MRP, and, where the app prints ratings, the share of the ratings shown on the three most-rated brands. Its signs of demand come from three: the share of stores read that carry the category, how many of its products a typical store lists (the app’s own range decision), and the ratings shown per rated product. A part an app does not give for most of its categories is left out for that app, never counted as zero. Each score is the mean of its parts, and open, contested and crowded (light, moderate and strong) are its thirds among the app’s categories.

Room to enter is a filter of ours, the same for every category: strong signs of demand on an open or contested shelf, or moderate signs on an open one. A category in depth shows its brands, what sells by the apps’ own marks (ratings, a “sold” badge), the price ladder (the price per 100 g, 100 ml or piece in eight equal bands; an empty price point is a band with no product, or under 3%, between two that have them), its pack sizes, the cities where its shelf is thin (at least 15 points under the app’s share across India, over 10 stores read), the plain keywords whose words name it, and the products first seen on our list at the latest read. First seen is not a launch: the October read reached deeper into every shelf than the reads before it. A rival’s range against yours is one app at a time: the categories where it has products and you have none, the ones you share, and its products first seen at the latest read. Listings are not sales, and nothing here is added across apps.

On Blinkit, Zepto, Swiggy Instamart and Flipkart Minutes every search we read carries the app’s own mark on a paid result, and we count every one: the brand, the keyword, the city and the position on the page. BigBasket and JioMart show a paid result only to a search that asks for it, the way their own pages ask. On BigBasket we asked on every keyword (through the app). On JioMart we asked on the head keywords, and its other keywords have blank ads cells. On JioMart a paid slot follows the category of the first result and not the word typed. Amazon Now shows sponsored results to signed-in shoppers only, and we read as a guest, so its paid columns are blank. A blank is not a zero. Everything here is a slot seen on a page. We do not see spend, bids, clicks or return, and no rupee figure is shown. We say a brand "held a paid slot on searches for a keyword": apps also place paid slots by category, so this says where the slot appeared and nothing about what the advertiser chose. "Cities with a paid slot" is the cities where a paid slot of the brand was seen among the results, out of the cities where the keywords it held one on were read; a paid banner is counted apart, with its own cities. The paid share of the first 10 is the part of the first ten positions marked as paid. "First result paid" is the share of searches whose very first result was paid.

On Swiggy Instamart each paid result states the keyword it was served on. We publish that keyword and one fact worked out from it: whether it is the word that was searched or a wider match. A brand that an app itself owns is marked as the app’s own label. Campaign ids, account ids and anything that looks like a cost in what an app returns are never published.

For a brand on one app we take the plain searches of its own categories (a search is the brand’s when its products are in the first 10, or when the app’s own category that fills the first 10 is one the brand sells in) whose paid slots were read, and leave out those where the brand held a paid slot in five or more cities. What is left is listed two ways. Where another brand, never a sister brand of the same company, held a paid slot in five or more cities on a first 10 that is a tenth or more paid, the search is listed with every such brand, widest first, and the brand’s own best unpaid rank there; the screen picks out the rivals the reader watches. Where paid slots were seen but fill under a tenth of the first 10, the search is listed as one with few paid slots.

A search with no paid slot at all is counted and never listed: an app may not sell slots on a search, or the law may keep ads off it, as with tobacco and infant formula. Nothing here says a slot would pay off or what it would cost; it is where slots appeared, read as a guest. The same paid slots are also added up by the kind of search each keyword is filed under in our list (snacks, beverages and so on), on a brand’s ads page and on its row beside its rivals.

For each brand name in our keyword list we count the brands that held a paid slot when that name was searched. A brand is named there only where it held a paid slot in five or more of the cities read; below that it is counted with "others". A sister brand of the same company is not listed as another brand, since a company advertising on its own label is not a rival on its name. This is the usual case and not an exception: on most brand names, another brand held a paid slot in five or more cities.

Amazon Now gives a product a title and no brand. On the Search screens we take its brand from the first words of the title, checked against the brands the other apps state for their products. A title that matches no known brand is left without one. We do not guess.

Where a product’s listings on different apps are matched as the same product and the same pack, the Products screen shows them as one row under one title, with each app’s listing beside it: the stores it was seen in, its median price over those stores, the MRP that app displays, and the share of looks that found it in stock. A product is "priced apart" when its median is 5% or more higher on one app than on another. "Lowest on" names an app only when its median is clearly below every other app that lists the product; when two apps are within 1% of each other, none is named. These are medians from different stores read at different times, so a difference is stated with both prices and not called a saving.

A listing we could not match with confidence is shown alone, and "no match found" on an app does not mean "not sold there". Your list keeps the products you choose in front of you, on every app where they are matched.

A unit price is the median price of a listing divided by everything in its pack: a pack of 2 x 500 g is priced over one kilogram. It is given per 100 g for weight, per 100 ml for volume and per piece for packs sold by count, and only where an app states the pack. A size we could only find in a product’s name is not used, because it is too often some other quantity: the protein in a bar is not the weight of the bar.

Beside a unit price we may show the range for its category on the same app: the price a tenth of the category’s products are under, the middle price, and the price a tenth are over. This needs at least 20 products seen in 20 or more stores, and at least half the category’s listings with a unit price. Categories are broad (cream is filed with milk), so the range gives context and is not a judgement. We do not compare two apps’ ranges, since each app files products its own way, and we show no range per piece: a piece of one thing is not a piece of another.

Product search looks through every product we read: one row per product, however many apps list it. A product is found when its title, its brand and its pack between them contain every word you typed, in any order. A word matches the start of a word, a number must match a whole number (1 kg does not find 1.5 kg), and "500ml" and "500 ml" are the same. Titles with the words together, in the order typed, come first; within each group the product seen in the most stores comes first.

Every signed-in account sees which apps list a product. Its price and the number of stores follow the plan: your own brand and the rivals you watch, and every brand on Pro. A product that no app listed in the stores we read is not in the search at all, which is not the same as not sold.

For each product we know the stores it was seen in on an app, and the stores where any product of its brand was seen on that app. "Least seen" puts a brand’s products in order of the first as a share of the second. It is shown on the apps where the app itself states the brand and the brand was seen in at least 50 stores. Unless you ask, it leaves out products seen in fewer than 20 stores, since those are more often a trial, a regional line or old stock than a gap.

Both numbers count sightings. Pages stop at a limit, and in our first reads most category shelves were read for their first few pages only, so a product low on a long shelf was not seen in a store even where it was listed. So "least seen" means "not listed, or low on the shelf". Both are worth a look, and it is not printed as a number of stores that do not sell the product. The claim that a product is missing from a store is made on Where you are missing, and only where a shelf was read to the end.

The Market and Categories screens count what the shelf read reached, one app at a time. A listing is one product in one store, as it was at our last look in the read. A brand’s share of shelf in a category is its listings divided by all the category’s listings on that app. "Held by the top 3 brands" is the same sum for the three largest. Nothing is added across apps: they differ in shelves, page limits and store counts, so two apps’ figures are shown next to each other and not summed.

Some figures are held back on purpose. A category needs at least 20 products on an app, seen in at least 30 of its stores, before that app gets any figure for it. Brand figures need at least 90% of the category’s listings to carry a brand the app states, because a listing with no brand counts for nobody; and a leader is named only where its lead is larger than all the listings with no brand. Own labels are the labels on our own list, so their share is a minimum. Two apps are compared on price only over products matched as the same product and pack on both, by counting which is more than 1% lower: a count, and not a saving. Store patterns (out of stock by kind of store) are our own reading of counts next to each other, and claim no cause. A blank on these screens is a figure we do not have or would not print. It is not a zero.

We keep several reads of the shelf, taken weeks apart. For each product on each app, a read gives one point, and only where the product was seen in at least 20 stores: its median price over everything we read, the MRP most listings showed, and, on the apps that publish stock, the share of looks that found it in stock. A product’s change runs from its first read with a point to its latest read with a point, and a move within 1% counts as level. A read with no point is a break in the line. It is not a zero.

The reads did not all go equally deep, because page limits changed between them. So the number of stores a product was seen in is not compared across reads, and prices are. While the latest read is still under way, and until it has covered 80% of an app’s stores, it is marked partial: its prices are those of the stores read so far, and no product is called "not seen in the latest read". That is never said on Swiggy Instamart or Flipkart Minutes. "Not seen" is not "delisted". The same-product basket for an app is the middle value of later price over earlier price across the products both reads hold. It has no weights and no sales in it, and it is not a consumer price index.

Shelf draws two kinds of map. The map of India puts one mark on each city, sized by the figure it stands for: the area of the mark follows the figure, so a city with four times the stores gets a mark four times the area, down to the smallest mark that can still be seen and pointed at. The mark is placed at the middle of a square about 9 km across that holds the middle of the city’s stores, on the apps whose stores we list. It stands for the city and not for any store, and a city we have no square for is listed and not drawn. The street map opens only when you ask for it, and draws each store as the 1 km square it is in. No latitude or longitude is published on either map.

Colour means the same on a map as everywhere else in Shelf. Green is you: on the shelf, in the first 10, in stock. Red is a rival where you are not, or your product out of stock. On a map of India a city turns red where a fifth or more of our looks found the product out of stock; that line is our own, and the legend states it. A ring is a place we read where there was nothing to count. Everything else is black. On a product’s street map coloured by price, green is below the price most stores show and red is above it, the same colours as a price cut and a price rise in What changed. A category’s street map splits one app’s stores into thirds by how many products of the category each lists; where most stores stop at the same page limit, only a store away from that count stands out. A map of sizes draws each category as a rectangle whose area is its number of products and whose shade is one more figure, named under it. No map here is a forecast, and none adds figures across apps.

A store page lists every product we read at that store, grouped by the category of the shelf it was read on, with the price and stock shown at that time. A product is named as the app lists it: name, brand and pack. A product seen on a shelf of a category other than its own is shown where it was read. The names of brands you do not watch are held back below Pro; the number of products on each shelf is complete on every paid plan.

For a category we also know, store by store, how many products its shelves listed. Where the app cut a shelf short, that count is a minimum. In our first reads most Blinkit shelves stopped at the same page limit, so many stores are level at the top and the order among them means nothing. Flipkart Minutes, Amazon Now and JioMart have no store page and no store list on any plan.

The Stores screen puts together two things we know about every dark store. One is the shelf read: the products and categories listed at the store, which is also where a brand was or was not seen. The other is a later check of each app at each store’s own location: the hours it states, the minutes it promised, the kinds of range the store ships, the wider ranges that reach it, and on Blinkit whether its print service answered. The two were read on different dates, and a store can change in between, so every screen that shows both prints both dates.

A store here is a store id as the app issues it, from our monthly list of stores plus any id the shelf read met that the list does not have. An app can have two ids at one site, so there can be more ids than buildings. A store can have its hours and reach and no shelf figures yet; a brand is only counted over stores with shelf figures. A store is placed by the 1 km square it is in. Flipkart Minutes, Amazon Now and JioMart are shown by city and state only, and their stores are not listed on any plan.

Each of these is one check of the app at the store’s own location, on the date shown. It is not an average and not a timetable. A store is "open 24 hours" when the hours the app stated for it cover the whole day; with shorter stated hours it "closes at night"; where the app stated no hours it is in neither group. Status is what the app answered at that minute: serving, paused, coming soon or not serviceable. Paused means it was not taking orders then, and does not mean the store has closed down. Promised minutes are what the app promised at the store’s own doorstep, its shortest trip, so the middle value across stores is not a typical delivery time and is not printed as one.

An app often serves one location from more than one store: a nearby express store for the everyday range, and a larger store further away for a wider range (long-tail), a still wider one (extended long-tail) or pharmacy stock, each with a longer promised time. A store’s roles are the kinds of range we saw it ship, in the answer for any location. A range "reaches" a store when it is offered at that store’s own location, whichever store ships it. So a store can be reached by pharmacy stock without holding any. On Zepto and Swiggy Instamart the same idea shows up as a second store serving the location.

Print stations and pharmacy stock are shown for Blinkit. On every other app those columns are blank, and a blank there is not a none. A print station is a store where the print service answered, including where it answered that it was down at that minute.

Products seen is the number of different products the shelf read saw at a store. Pages stop at a limit, so it is a minimum of what was listed and not the size of the store’s shelf. The range band is our own split, and labelled as ours: the stores of an app whose shelf lists came back almost in full are put in order of products seen and cut into thirds: compact, standard and wide. It ranks what the pages listed. It is not floor area and not the store’s true range, and it is not shown on apps whose pages cannot rank stores, which today are Zepto and Swiggy Instamart.

A store carries a category when a shelf of that category listed at least one product there at the shelf read. "Stores carrying it" is counted out of the stores where a shelf of the category gave an answer. "Products a typical store lists" is the middle value, over the stores that carry it, of the products listed on the category’s shelves: it shows how crowded the shelf is where a brand competes. Both are minimums, because pages stop at a limit. No row for an app means we have no shelf of that category for it. It does not mean the app does not sell it.

For each app we count the stores where at least one product of a brand was read, and show that beside two totals: the stores of the app where the read got any answer, and every store of the app on our list. A share is always over stores read. The same counts are given by zone, by state, by kind of city and by city. A city is counted in the state most of its stores are in, so a state is the sum of its cities and a zone the sum of its states. A zone is one of the six groups of states used by the Zonal Councils and the North Eastern Council. Two brands shown together are compared over the same stores read. This counts where a brand was seen. It does not say where a brand is not sold: every app cuts its lists short somewhere, so a store with no sighting is not a store without the brand, and on Swiggy Instamart and Flipkart Minutes every count is a minimum. A brand is counted only where the app itself names it on the listing. For a category the count is the stores where at least one product of the category was read, with the products and brands read there, how many products a typical store lists, the share in stock at the last read and the middle price.

Some apps print on a product how much it is bought: the number of ratings it has, and on BigBasket a badge such as "150+ sold (1 mo)". Within a category we rank an app’s products by those marks, and its brands by their share of the ratings shown. They are the apps’ own figures, as printed the last time we read one. We do not measure sales and do not estimate them. A count of ratings keeps growing for as long as a product is listed, so it favours older products: it gives an order, and is not a rate of sale. BigBasket’s badge was on most of its products in August and on few since, so every badge keeps the day it was last printed: one printed in the read on the screen ranks a product, and an older one is shown with its day and ranks nothing. Amazon Now names no brand on a listing, so its products are ranked and it has no table of brands. A mark on one app is not compared with a mark on another, and a category needs ten products with a mark on an app before it is ranked there.

For each kind of store (open 24 hours or closing at night, reached by pharmacy stock or not, wide or compact in range) we count the stores of that kind whose shelves were read and, of those, the stores where the brand was seen. It counts where a brand was seen. A store where the brand was not seen is not called a store without it on this screen: that needs a shelf read to the end, and it is worked out store by store on Where you are missing. On Swiggy Instamart every such count is a minimum, so two of them are not compared.

"Thinner here" is our own reading of that table: a kind of store where the brand was seen in a clearly smaller share of stores than in the opposite kind, where both groups have at least 30 stores and the shares are at least 10 points apart. It states two counts and no cause. Kinds of store overlap: stores that close at night are also, more often, the ones with a compact range.

Every store on the apps we read is read again every day, JioMart and Amazon Now twice a day. The largest apps take up to two days to read from end to end, so every figure rests on one whole read of its app. Every read is kept, and that is what lets a product be followed store by store: when it went out of stock, how long it stayed out, and how fast its stock runs down, shown as bands, never as units. Which reads an account sees follows its plan: Pro sees every day’s read, and a read every hour in the target cities it names with us; Free is refreshed every week.

Figures published while a read is still under way cover only part of it, and are marked "partial read" wherever they appear. They do not replace the figures of a finished read. New figures are published when a read is finished.

The glossary lists all 206 terms in alphabetical order, each in the one sentence the screens show beside it, with a link back to its rule on this page.

1. 1If a figure about your brand looks wrong, write to trust@seetheshelf.com.
2. 2We aim to answer within one working day.
3. 3We read the store again, and print the new reading next to the original.
4. 4If we were wrong, we change the method and not only the figure, and the changelog says so that day.

Corrections asked for and read again so far are listed on the changelog, with their dates.