AI visibility tracking: how to measure whether AI answers mention you

Rank tracking assumes a stable, shared results page you can sample on a schedule. AI answers have none of those properties, which is why position number one has no meaning here and why no tool has credibly recreated it.

The strongest evidence available is a Productrise study run between 9 and 31 August 2026 across the US and UK, covering 2M+ listings and 100k+ SERPs. It found 1.28% product overlap between classic Google search and AI Mode. On the small set of products that did appear in both, the main seller differed 49.6% of the time.

The same study found AI Mode prices 21.6% higher on matched products, with a median AI Mode price of $149 against $100 in classic results. Read plainly, AI Mode is selecting a different set of products and a different set of sellers, on different criteria.

So a rank tracker built on classic results is measuring a neighboring channel, not this one. Three further properties break the model outright.

• Answers vary by session. The same prompt run twice an hour apart can return different brands, with no change on your side.
• Answers vary by account. Memory, chat history and personalization mean two users asking identically can see different products.
• There is no position to record. You are mentioned or you are not. Being third in a paragraph is not a rank, and treating it as one invents precision.
• Demand is invisible. No platform publishes how many people asked your query, so you cannot weight a prompt by how much it matters.

Manual work is the consequence of that last point specifically. The problem is not a missing tool, it is a missing sampling frame. This sits inside the broader shift to agentic commerce, where the reader of your page is often a program.

Four things are genuinely measurable right now. Everything beyond them is estimation, and it is worth knowing which side of that line any number you report sits on.

Mention presence on a fixed prompt set

Given the same ten prompts on the same surfaces, you can record whether your brand or product was named. This is binary, repeatable and the foundation of everything else on this page.

Citation presence in the linked sources

Where a surface shows its sources, you can record whether your domain appears among them. Citation and mention are different events. You can be cited as a source and never named as a recommendation, and the reverse happens too.

Referral traffic from AI surfaces

Your analytics will show some referrals from chat and AI search hostnames. Treat this as a floor rather than a count. In-answer purchases, wallet checkouts and copied product names produce zero referrer data while still being caused by AI visibility.

Product-level appearance for your own SKUs

Ask an agent for a product you sell, by name, and record whether it is found and whether the price and variant returned are correct. This is the most diagnostic of the four, because a failure here is a data problem you can fix rather than a ranking problem you can only influence.

What you cannot measure

• Impressions. Nobody publishes how many answers were generated for a query.
• Query demand. There is no keyword volume equivalent for prompts.
• Full competitive share. You can only compare within your own prompt set.
• Causality. A change in mentions can come from your fix or from a model update the same week.

A first baseline takes about two hours and needs a spreadsheet, nothing else. Do it before you change anything, because a measurement taken after the fixes is not a baseline.

1. Write down your ten highest-intent buying queries. Use the phrasing a shopper would use out loud, not your keyword list. "Best waterproof hiking boots for wide feet under $180" rather than "waterproof hiking boots".
2. Pick your surfaces and freeze them. ChatGPT and Google AI Mode at minimum. Add Perplexity, Copilot or an agent you have access to if your customers use them. Changing the list later breaks comparability.
3. Run every prompt in a clean, logged-out session. Personalization from your own account history will flatter you. Use a fresh browser profile or incognito for every run.
4. Run each prompt three times. Record all three. Where they disagree, the majority result is your reading and the disagreement itself is worth noting.
5. Record the fields, not the vibe. Date, surface, prompt ID, mentioned yes or no, competitors named in order, your domain cited yes or no, and the price quoted if a product appeared.
6. Paste the raw answer text into the sheet. You will want the exact wording in three months when you are trying to work out what changed.
7. Ask for five of your own products by name. Note whether each is found, and whether the price and variant returned match your live listing.
8. Segment AI referral hostnames in analytics and annotate today's date. This gives you a floor number to read alongside the prompt set later.

Scoring is simple. If step 7 fails, your product data is the problem, and fixing presence before fixing data just gets you found and then misquoted. If step 7 passes and step 4 shows you absent, the work is visibility rather than data. Both paths are covered in LLM SEO.

The prompt set is the measuring instrument, so its design decides whether your numbers mean anything. Ten prompts is enough to see movement and small enough that someone will actually run it every week.

How to choose the ten

Pick for purchase intent, not for volume. A prompt someone types with a credit card open is worth more to you than a popular research question you will never convert. Your sales team's most common inbound questions are a better source than a keyword tool.

Spread them across shapes so one model quirk cannot swing the whole set. A workable mix is three category plus constraint prompts, two comparison prompts, two problem-first prompts where the shopper describes a situation rather than a product, two budget-bounded prompts, and one brand alternative prompt.

Rules that keep the set comparable

• Freeze the wording. Store the exact string with an ID. A reworded prompt is a new prompt and resets that row's history.
• Keep your brand name out of at least seven of the ten. Prompts containing your brand measure recall, not discovery, and they will make the set look healthier than it is.
• Write in a shopper's voice. Full sentences, casual phrasing, the odd typo left in if that is how people type.
• Include one prompt you expect to lose. A set you always win tells you nothing and quietly drifts into vanity.
• Review the set quarterly, not weekly. Changes break the time series, so batch them and note the date you changed them.

One practical warning from running these: prompts that name a price cap tend to be the most volatile, because model price data is often stale. If a prompt swings wildly week to week, check whether the cause is your visibility or a wrong price being used to filter you out.

Weekly is the right cadence for most merchants, and monthly is acceptable if weekly means it does not happen. Daily is not useful, because session variance is larger than any real week-to-week movement you would act on.

At a glance: the fields to log every run

• Date and time, because surfaces ship changes mid-week and you will want to line events up.
• Surface and prompt ID, never the prompt text retyped, which invites silent edits.
• Mentioned, yes or no, as a strict binary with no partial credit.
• Competitors named, in the order they appeared, which is the most useful column you will keep.
• Your domain cited, yes or no, logged separately from mention.
• Price and variant quoted, where a specific product was returned.
• Raw answer text, pasted whole, because summaries lose the detail you will need later.

Share of answer, and where the number is soft

Share of answer is the percentage of prompt and surface combinations in which you were mentioned. Ten prompts on three surfaces gives thirty slots, and being named in nine of them is 30%. It is easy to compute and easy to over-read.

Three things make the number soft. Your sample is ten prompts out of an unbounded question space. Each prompt counts equally even though real demand behind them is wildly unequal and unknowable. Session variance means the same week rerun can move the figure by several points with nothing having changed.

So report it as a band and a direction, not a decimal. "Roughly a third of our tracked prompts, up from about a fifth in July" is honest. "31.7% share of answer" is not, and the second version will be quoted back at you in a board deck.

AI visibility tools are real and improving, and they solve the labor problem rather than the measurement problem. Knowing exactly which half they solve keeps you from over-trusting the output.

What they do well

• They run prompts at scale. Hundreds of prompts across several surfaces on a schedule, which no human will sustain manually.
• They log consistently. Structured storage of answers and mentions beats a spreadsheet someone updates when they remember.
• They track cited domains. Seeing which sources a surface keeps returning to is genuinely actionable for content planning.

What they cannot do

• They cannot see demand. No vendor has access to how many real people asked anything, so every share figure is unweighted.
• They sample their prompt list, not your customers. A vendor's default set will not match your buyers unless you replace it, and most teams never do.
• They run clean sessions. Real users carry history and memory, which changes what they see.
• They stop at the answer. Whether an agent then checked out with the correct SKU is outside what they observe.

Practical position: buy a tool once your manual set is running and you have outgrown it, not before. The tool scales a method you already trust. It will not tell you which prompts matter.

Who should own this internally

Give it to whoever owns organic content, with a standing dependency on whoever owns the product catalog. The content owner runs the set and reads the trend. The catalog owner fixes what the product-level checks expose, because most misses trace back to data rather than to writing.

Budget two to four hours a month once the sheet exists. It does not need a new hire, and handing it to an agency without your own copy of the log is how the history gets lost at renewal.

Four things take under a week and make the measurement trustworthy. The harder work after that is what actually moves the number.

Easy fixes

Freeze the prompt set in a shared sheet with IDs. What good looks like: anyone on the team can run the set without asking what the prompts are, and no one edits a string without noting the date.

Add three repeats per prompt. What good looks like: every logged reading is a majority of three, so a single odd answer no longer moves your reported figure.

Segment AI referral hostnames in analytics. What good looks like: a named segment and an annotation on the date you started, so the floor number has a beginning.

Put a name and a recurring calendar block on it. What good looks like: the run happens on the same weekday whether or not anyone chases it.

The harder work

Fix the product data the checks expose. Variant grouping, precise taxonomy, literal product fields and real-time price and stock are what Shopify's agentic-ready product data guidance names as the requirements, and Shopify reports AI-referred orders growing nearly 13x year over year alongside AI-referred visitors converting at nearly 50% higher rates than organic search.

Then earn citations rather than chase mentions. The sources a surface returns to repeatedly are the ones with specific, checkable, well-structured information. That is a months-long content position, not a sprint, and it is covered in how to rank in ChatGPT.

Last, extend the log to agent behavior as well as answers. Checking whether an agent returns the correct SKU and price for your products is a different test from whether you were mentioned, and it catches the failures described in AI shopping agents.

QuickAds does not sell an AI visibility dashboard. We treat the prompt set as a diagnostic and then work on the three things it usually exposes. In our own runs, the most common reason a merchant is absent from an answer is not weak content. It is a catalog an agent cannot read cleanly enough to quote.

Shopify catalog fixes

Variant grouping, taxonomy cleanup, rewriting product fields into literal machine-readable language, and live price and stock sync. When a product-level check fails because an agent cannot find a SKU you definitely sell, this is the fix, and it is the one that changes the measurement fastest.

Catalog video with correct tagging

Product video generated per SKU with structured metadata and schema attached, so the information inside a video is parsable rather than invisible to the surfaces you are tracking.

Performance creative

100+ creatives per month at a 5-7 day turnaround, with creative intelligence trained on 32M+ ads. Relevant because AI surfaces now carry paid placements beside organic recommendations, so both sides of the screen need assets.

The coverage argument matters here. We run creative intelligence, strategy, production and campaign management as one chain. Analytics-only tools report and stop. Premium production shops build assets and never touch your product data. Visibility problems usually live in the seam between those two.

Being clear about limits: we cannot promise you a position in an AI answer, and neither can anyone selling you one. The surfaces change monthly, and a model update can move your numbers in a week you shipped nothing. What we commit to is throughput on the work itself, at software from $299 per month or managed from $5,000 per month.

If you want the diagnostic run against your own catalog before you commit to anything, the free ad account audit is where to start. Bring your ten prompts and we will run them with you.