Share of search is the percentage of the results on a retailer search page that your products occupy for a given keyword, tracked by term, position, and retailer, and usually split into organic and sponsored placements. It is the headline metric of every digital shelf analytics platform, and it is genuinely useful.
It is also a measurement of a web page, not a measurement of a person. That distinction is the entire subject of this article.
The short answer
Share of search proves one thing well: whether your products were present and prominent on a results page, at a moment, from a particular vantage point. It cannot prove that a shopper saw that page, that the ranking reflects preference, or that a position gain caused a sales gain. Those three claims are the ones most digital shelf decks actually make.
The fix is not to abandon the metric. It is to treat share of search as an exposure estimate and to pair it with direct shopper evidence that answers the question a scrape structurally cannot: among the people who searched that term, what were they trying to do, and why did they choose what they chose?
What a scrape actually captures
A digital shelf platform sends an automated request to a retailer, receives a results page, and records which products appeared where. Every number downstream inherits the properties of that request.
| The scrape | The shopper |
|---|---|
| Logged out, or logged into a monitoring account | Logged in, with purchase history |
| One location, usually a data-centre IP | A real delivery postcode with real local inventory |
| One moment, on a schedule | An unpredictable moment, mid-mission |
| One phrasing of the keyword | Whatever they actually typed, including misspellings |
| Sees the full grid | Sees a viewport, and mostly the top of it |
| No intent | An intent that determines which result counts as a win |
None of these are defects in the tooling. They are the price of automated measurement, and every vendor in the category faces them equally. The error is forgetting they exist when the number reaches a slide.
Gap 1: the page you scraped is not the page they saw
Retailer search results are personalised, inventory-filtered, and frequently experimental. A shopper with a two-year history of buying your competitor sees a different grid from a shopper who has never bought the category. A shopper whose local store is out of your pack size may not see it at all, or may see it demoted. And retailers run continuous ranking experiments, which means some fraction of shoppers are seeing a treatment page that no scrape is sampling deliberately.
The consequence is subtle and worth stating plainly: share of search has a denominator problem. You are reporting a percentage of a results page that does not correspond to any particular shopper's experience. It is closer to a weather station reading than to a census.
This is the same class of problem we described for prices in personalized pricing research, where a single posted price stops existing and every method that assumed one quietly returns a confident average of incompatible experiences. Assortment visibility is now going the same way, a few years behind.
What to do about it: stop treating a single share-of-search number as a fact and start treating it as a distribution you have sampled badly. Scrape from multiple postcodes if your tool allows it. Report a range. And validate the range against people, which is the subject of the last section.
Gap 2: position is confounded with paid placement
The top of a retailer results page is increasingly for sale, and the scale of that market is now visible in public filings rather than industry estimates.
In its FY2025 Form 10-K, Amazon reported advertising services revenue of $68.6 billion for 2025, up from $56.2 billion in 2024 - roughly a 22% increase in a single year. For scale, third-party seller services brought in $172.2 billion and consolidated net sales were $716.9 billion. Advertising is no longer a side business attached to the shelf. It is a substantial share of what the shelf is for.
That matters for measurement in a specific way. When your share of search rises, at least four different things may have happened:
- Your organic relevance improved.
- You, or a reseller, increased sponsored spend on that term.
- A competitor cut spend on that term.
- The retailer changed how many ad slots the page carries.
Only the first is a marketing result you can bank. The other three are auction dynamics and retailer product decisions, and they can move your number substantially without anything changing about your product or your shoppers.
The Federal Trade Commission and 17 state attorneys general, in the monopolisation complaint filed against Amazon on 26 September 2023, put the concern in unusually direct language. The FTC alleges Amazon is "degrading the customer experience by replacing relevant, organic search results with paid advertisements - and deliberately increasing junk ads that worsen search quality and frustrate both shoppers seeking products and sellers who are promised a return on their advertising purchase." The complaint further alleges Amazon "can bury discounting sellers so far down in Amazon's search results that they become effectively invisible," and that combined fees "force many sellers to pay close to 50% of their total revenues to Amazon."
These are allegations in litigation that Amazon disputes, and nothing here should be read as a finding of fact. But they are allegations about the measurement environment made by the agency with subpoena power over it, and that alone is a reason to stop reading rank as a neutral signal of merit.
If you want the paid-versus-organic problem handled properly on the media side, we covered the attribution half of it in retail media measurement. This article is about the organic half, which gets far less scrutiny because nobody sends you an invoice for it.
What to do about it: never report blended share of search. Split organic and sponsored, always, and report the sponsored line next to the spend that produced it. A blended number that rose because you bought it is not a finding.
Gap 3: rank tells you where you appeared, never why they chose
This is the gap that no amount of better scraping closes, because the missing variable was never on the page.
Suppose your share of search on a category term goes from 12% to 19% and sales are flat. There are at least six explanations, and the scrape distinguishes none of them:
- The extra visibility landed on a keyword with browsing intent, not buying intent.
- Shoppers saw you, considered you, and rejected you on price, format, or reviews.
- Shoppers saw you and did not recognise the brand, so the impression did nothing.
- The additional placements were below the fold on the devices most of your shoppers use.
- Your gain came at the expense of your own other listing, not a competitor.
- Shoppers clicked, read the detail page, and left because the information they needed was not there.
Every one of those implies a different action, and three of them mean spending more on visibility would be actively wasteful. This is the classic gap between what behavioural data records and what it explains, which our documentation covers in attitudinal vs behavioral research and stated vs revealed preferences.
The version of this that costs the most money is the fifth one. Self-cannibalisation is invisible in share of search by construction, because your own listings taking each other's clicks still counts as your share going up.
How to close the gaps with shopper research
The practical answer is triangulation: keep the scrape for coverage and trend, and add a small, fast, repeatable study that supplies the variable the scrape cannot observe. Our guide to triangulation in research covers the general principle; here is the specific design for the digital shelf.
Recruit shoppers who bought in the category in the last 30 days, in the retailers you care about, and run a study that mixes conversational depth with countable structure. Using the six structured question types, a workable instrument looks like this:
| What you need to learn | Question type | Example |
|---|---|---|
| The words they actually search | open_ended | Tell me what you typed the last time you bought this |
| Whether they scrolled at all | yes_no | Did you look past the first row of results? |
| How the choice was made | ranking | Rank these five factors by how much they drove your pick |
| Which brands were even in play | multiple_choice | Which of these did you consider? |
| The retailer they defaulted to | single_choice | Where did you buy it? |
| Strength of the preference | scale | How likely were you to switch if your first choice was unavailable? |
Two design notes that matter more than they look.
Ask for the query verbatim, not the category. Teams consistently discover that their tracked keyword list is a marketing artefact. Shoppers search for the problem, the occasion, or a competitor's brand name as a generic, and the term you have been optimising is one nobody types.
Ask what they did not choose and why. The rejected set is the single most valuable output of digital shelf research and it appears in no scrape, no clickstream, and no POS file. It is the shelf equivalent of the point made in market basket analysis: a co-occurrence score describes what happened, and several incompatible shopper realities produce the same score.
Where Koji fits
The reason digital shelf teams historically stopped at the scrape is not that they thought rank explained behaviour. It is that the alternative - recruiting category buyers, moderating interviews, coding transcripts, writing a readout - took six weeks and a research headcount, by which point the range review had happened.
Koji removes that constraint. You write a brief, Koji's AI consultant turns it into a discussion guide, AI-moderated voice interviews run with real shoppers around the clock, and thematic analysis lands as a one-click report. Every interview asks the same questions with the same follow-up logic, so segment differences are not artefacts of a moderator finding one group more interesting than another - the problem we describe in interviewer bias. Structured question types run alongside the conversation, so you get countable data and quotes from the same session.
That turns "why did our share of search gain not convert" from a quarterly investigation into a question you answer this week, for the cost of a fraction of one month of digital shelf tooling. No research expertise required, and the output arrives in hours rather than weeks.
If your share of search is up and your sales are not, the answer is in the heads of people your dashboard never talked to. Start a free study on Koji and have the rejected set by Friday.
Related reading
- AI Search Visibility (2026)
- Winning the Buy Box (2026) - why ranking data records the algorithm's decisions, not your customer's preferences.
- Agentic Commerce Research (2026) - what happens to the digital shelf when an AI agent does the looking.
- Retail Media Measurement - the paid half of the same page.
- Customer Research for Ecommerce Brands - the wider playbook.
- AI Research for E-Commerce - how Koji runs ecommerce studies.
- Ecommerce Customer Research Guide - methods and question sets by journey stage.
- Product Detail Page Research (2026) - the page they land on after the grid, and what a PDP test cannot see.
Frequently Asked Questions
What is share of search on the digital shelf?
It is the percentage of results on a retailer's search page that your products occupy for a given keyword, measured by term, position, and retailer, and normally reported separately for organic and sponsored placements. It is an exposure metric: it tells you whether you were present and prominent on a page, not whether a shopper saw it or acted on it.
Is share of search the same as share of voice?
No. Share of voice is usually a media metric covering advertising presence across channels. Share of search on the digital shelf is specific to retailer and marketplace search results pages, and it mixes organic ranking with paid placement unless you deliberately split them. Reporting them blended is the most common error in the category.
Why does my share of search rise while sales stay flat?
The common causes are that the gain landed on low-intent keywords, that it came from sponsored placements you paid for rather than organic relevance, that the extra placements fell below the fold, that shoppers saw you and rejected you on price or format, or that your own listings cannibalised each other. A scrape cannot distinguish these, which is why the diagnosis requires asking shoppers directly.
Can digital shelf analytics tools measure personalised results?
Only partially. Most tools scrape logged out or from monitoring accounts, from a limited set of locations, on a schedule. Real shoppers see results conditioned on purchase history, local inventory, device, and any ranking experiment they have been assigned to. Treat a share-of-search figure as a sample from a distribution you have not fully observed, and report ranges rather than single points where your tool supports it.
How many shoppers do I need to interview to explain a share of search change?
For a diagnostic study on a single category and retailer, teams typically reach stable themes in the range of 15 to 30 interviews, with more needed if you are comparing segments or retailers. Our guide on how many interviews are enough walks through how to decide rather than guess, and the answer depends far more on how many distinct shopper situations you are covering than on the size of the market.
How is this different from just reading reviews and search suggestions?
Reviews and search suggestions are real signals, but both are heavily filtered. Reviews over-represent very positive and very negative experiences and the people willing to write, and suggestion data reflects the retailer's own query normalisation. Neither contains the rejected set - the shoppers who considered you and bought something else - which is the specific output that changes a shelf decision.
