Free AI visibility check for Shopify stores

What changed

Across your tracked buying questions, AI engines recommended you in 3 of the last 5 decisive scans, up from 2 the week before. Engines began naming you for “best home espresso machine under $500”: recommended in 3 of 5 scans, up from 0.

Why it likely changed

For that question, engines started citing two review sites that list your machine. On “quiet coffee grinder for small kitchens” nothing moved: a competitor is still recommended in 4 of 5 scans.

Do these 3 things

1. Close the grinder gap. Engines cite a handful of review sites when answering that question; getting your grinder reviewed there is the most direct route in.
2. Fix 2 feed issues. Two products are missing a valid GTIN. Add the barcode in Shopify admin.
3. Add the comparison your buyers ask. Track a head-to-head question your catalog can win.

Made-up data for a fictional store. In a real report every number is a count from your own scans (recommended in N of the last M decisive scans). “Unsure” answers are left out, never counted as absence.

Every claim on this page maps to shipped, tested code. When the methodology changes, this page changes in the same release. Methodology changes never silently rewrite history.

What we measure

ShelfSight asks the same buying-intent questions your customers ask, across ChatGPT (OpenAI), Perplexity, Gemini, Claude, and Google AI Overviews, and records whether the answers mention or cite you or named competitors.

Which engines a scan covers depends on where it runs. Paid plans scan all five every week, within a monthly engine budget (half the plan price on Starter and Growth, three fifths of it on Pro); if a month's budget runs out, scans pause and resume on the 1st. The free first scan in the app covers as many engines as its budget allows. The free public report on this site drafts 3 buying questions from your public product feed, asks each once on Perplexity, Gemini and Google AI Overviews, counts the decisive answers that recommend you, and checks the feed.

How a "mention" is decided

Deterministic rules, strongest signal first:

1. Citation domain. An answer cites a URL on your domain → counted as mentioned. This is the strongest signal and is never overridden.
2. Name match. Your brand (or a configured alias) appears as a whole word/phrase in the answer → mentioned. If your brand name is also an everyday word (e.g. a brand called "On"), a bare text match is never counted as a mention without citation corroboration; it's marked unsure instead.
3. Product-title overlap. Distinctive words from your product titles appearing together are at most unsure, never a confirmed mention. Words the question itself used don't count: an answer to "best alpine touring skis" says "alpine touring skis" whoever sells them.

Anything that doesn't clear these bars is not mentioned. We do not use fuzzy guessing; "smart wool" does not match "Smartwool" unless you configure it as an alias.

Why "unsure" exists, and why it never counts against you

AI answers are messy. When we can't decide honestly, we say so: unsure results are left out of the decisive count. They are never silently treated as absence (or presence). An engine that returns an empty answer is recorded as unsure, because no answer is not evidence you're invisible.

"Referenced" is a weaker claim than "mentioned"

The per-product view on your dashboard shows where engine answers reference individual products from your catalog. A reference means distinctive words from a product's title, beyond the words of the question, appeared together in an answer: the same conservative rule 3 above. It is token-level evidence, deliberately labeled referenced: weaker than a confirmed mention, and never a recommendation count. We show it because knowing which products surface (and which never do) is actionable; we label it honestly because the evidence is circumstantial.

Stability counts

A single AI answer is one sample; engines vary run to run. So we lead with counts: "recommended in 3 of your last 5 scans", per question and engine, with the answers behind each count shown.

• Your first scan asks each question 3 times on each engine, so the first count already rests on more than one answer.
• Ongoing scheduled scans are single runs; stability is computed over the rolling window of your recent scans. A time series is repeated sampling.
• When there's no decisive data, your dashboard says "no decisive scans yet", never a made-up zero.
• We never turn a count into a 0 to 100 score, a percentage of answers or a points change. Every number is a count, shown with the answers it comes from.
• Your competitors are counted on the same answers by the same matcher. Each store's count leaves out the answers that were unclear for that store.
• The comparison with other tracked stores shows two counts of decisive answers from the last 30 days, and appears only once at least 5 stores and 50 decisive answers are in. Stores track different questions, so it is context, not a ranking.
• Every result is stored with the engine's model version and timestamp. When engines change models, your history stays exactly as it was measured.
• Claude answers from 29 September 2026 come from Claude Sonnet 5; earlier ones came from Claude Opus 4.8. Two models answer differently, so Claude counts from before and after that date aren't like-for-like.

How the 14-day view and the weekly counts work

The 14-day view shows, per day, how many decisive answers there were and how many of them recommended you; unsure answers are left out, exactly as in your counts. A day with no decisive answers shows no bar, not a zero. The weekly view puts the last 7 days next to the 7 days before, as a count for each week. We never turn the two into a percentage or a points change, and a week with no decisive answers says so.

How drop alerts work

On Growth and Pro, each day we compare, for every prompt on every engine, your last 5 decisive answers with the 5 before them. The two sets never share an answer, and unsure answers are left out, exactly as in your counts. You get an alert when the newer 5 recommend you fewer times than the older 5: at least 2 fewer on high sensitivity, 3 on standard (the default), 4 on low.

• A prompt needs 10 decisive answers on an engine before it can alert. The first scan gives 3 and each weekly scan adds one, so alerts start after about 7 weeks of scans, later if some answers are unsure. Until then there is nothing fair to compare.
• An alert you dismiss doesn't come back until 5 new decisive answers are in.

Orders AI assistants sent you

This is optional and off until you turn it on in the app, which asks Shopify for permission to read your orders. The option only appears once Shopify has approved ShelfSight's access to order data, and even then nothing is read until you turn it on. We then read your orders from the last 60 days, up to the 2,500 most recent (if there were more, we say the older ones weren't read), and look at the visits Shopify recorded before each one.

• An order counts as coming from an AI assistant only on a named signal: the last or first visit came from the assistant's website (chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com, claude.ai and a few others), carried the tag the assistant adds (ChatGPT adds utm_source=chatgpt.com), Shopify labelled the visit's source as the assistant, or the order was placed through the assistant's own checkout.
• Orders with no referrer and no tag are counted as can't tell, never as AI and never as not AI. Some AI visits arrive that way (links opened from apps, copied links), so the AI count is a floor.
• Google AI Overviews and AI Mode send visitors that look exactly like Google Search, so those visits are never in this count. An order placed through Google's own checkout inside AI Mode or Gemini counts under Gemini when Shopify names that checkout on the order.
• We show counts next to every order in the window, never a percentage of your revenue, and never a claim that an assistant caused a sale.
• We keep, per order, only its ID, date, the assistant and the website or tag that named it, for 60 days. No customer details. Turning the feature off deletes it.

What we deliberately don't sell

• llms.txt. Google says its Search doesn't use these files, and a 2026 server-log study of 137,210 domains (Ahrefs) found 97% of the llms.txt files got no requests at all in a month. We don't generate one or charge for it.
• "AI schema". Google says no special markup is needed for its AI features, and Shopify themes already publish product structured data with price and availability. A second copy adds nothing we can measure.
• Article autopilots and silent bulk rewrites. AI answers lean heavily on third-party sources, and rewriting your pages without your review risks the content you care about.

What we check instead: whether your product data (category, identifiers, options, price and stock) is complete for the feeds AI shopping surfaces read, and whether your robots.txt lets the crawlers behind AI search (OAI-SearchBot, PerplexityBot, Claude-SearchBot, Googlebot, Bingbot) read your product pages. Blocking training-only crawlers such as GPTBot doesn't change whether assistants can cite you, and we say so. Google-Extended is different: it doesn't affect Google Search, but Google says it controls whether Gemini Apps use your pages to ground answers, so we flag it on its own.

An honest caveat about APIs

We query engines through their public APIs with web search/grounding enabled. API responses are a statistical proxy for what consumers see in the apps: close, current, and reproducible, but not pixel-identical. This is standard practice across the AI-visibility industry; we state it plainly because the alternative (pretending otherwise) would be dishonest.

The feed audit has no proxy problem: it checks your actual product data with fixed rules, against the published product feed requirements of the Agentic Commerce Protocol and against what Shopify Catalog uses to place and match a product (a standard product category, and option values on every variant). Every finding shows the exact rule and the exact fix.

How the weekly report is written

An AI model writes the weekly report from your scan and audit data, instructed to use only the counts we measured, to compare weeks only as counts (never as a percentage or a score), not to invent numbers or causes, and to say "no decisive data yet" when that is the truth; tests check those instructions.