AI visibility tracking — how to measure citations

Traditional SEO measurement had one enormous advantage: a stable, ordinal answer. You ranked fourth. Next month you ranked second. Everyone understood what happened.

AI answers give you none of that. Ask the same engine the same question twice and you may get different sources, different phrasing, and a different list of companies named. The temptation is to conclude the channel is unmeasurable and skip it. That is wrong, but the correction isn't to look for a position number that doesn't exist — it's to change what you count.

What to track

Five things, in rough order of how much they should influence decisions.

Two supporting signals sit underneath these. Crawler access: whether the engines' documented user agents are reaching your pages, verified in server logs. Referral traffic: visits arriving from AI products, which are small for almost everyone right now. Treating that traffic as the primary KPI will make you abandon the work before it compounds. Mention and citation rates are the leading indicators; traffic is the lagging one.

Building a manual baseline

You can do this without spending anything, and you should do it once even if you later buy a tool, because it teaches you what the data actually looks like.

1. Write the prompt set. Twenty to fifty questions, written the way a real buyer would type them into an assistant — full sentences, not keywords. Include four categories:

• Category questions — "what's the best way to automate invoice processing for a small business"
• Comparison questions — "how do AI phone answering services compare"
• Problem questions — "my team spends hours a week on manual data entry, what are the options"
• Brand questions — "what does DoubleTime AI do" and "is DoubleTime AI any good"

Brand questions are the ones most people skip and the ones that surface accuracy problems.

2. Fix the conditions. Use logged-out sessions or a clean profile so personalization and memory don't contaminate results. Record which engine, which mode, and the date. Run the same set the same way every time — comparability is the entire point.

3. Record structured results. One row per prompt per engine per run: mentioned (yes/no), cited with link (yes/no), position in the answer, competitors named, accuracy note. A spreadsheet is fine. Resist recording a subjective quality score; it won't survive comparison to last month.

4. Repeat monthly and read the trend, not the run. Weekly is noise; quarterly is too slow to catch a problem. A single run tells you almost nothing because of non-determinism. Three runs tell you a direction. The value shows up around month four, which is exactly when most people stop.

What the platforms give you for free

Two sources of first-party data are worth wiring up before you evaluate any vendor.

Google Search Console. In June 2026 Google introduced Search Generative AI performance reports in Search Console, giving site owners a view dedicated to visibility from generative AI features rather than only blended into overall Search performance. It rolled out to a subset of sites, so check whether yours has it. This is the only first-party, non-sampled data any engine currently publishes about AI-surface visibility — if it's available to you, it outranks every third-party estimate.

Your own server logs. Every major AI company documents its crawlers and publishes verification methods. Perplexity documents PerplexityBot for indexing and Perplexity-User for live user-initiated fetches, with IP lists published for verifying both. OpenAI documents GPTBot for model training, OAI-SearchBot for surfacing sites in ChatGPT search, and ChatGPT-User for user-initiated actions. Filter your logs for these agents, verify against the published IP ranges — user agent strings are trivially spoofed — and answer the eligibility question directly: are these systems reaching my pages at all, and which ones?

This is the cheapest diagnostic in the discipline and it is routinely skipped. A site that blanket-blocked AI crawlers out of training concerns will show zero citations forever, and no amount of content work changes that until the robots.txt and WAF rules are fixed.

What categories of tooling exist

The vendor landscape here is young, crowded, and turning over fast, so it's more useful to understand the categories than to memorize names.

• Prompt monitoring platforms. Run a defined prompt set against multiple engines on a schedule and report mention rate, citation rate and share of voice. The dominant category — essentially your manual spreadsheet, automated.
• AI modules inside established SEO suites. The major SEO platforms have added AI-visibility reporting to existing subscriptions. Often the pragmatic choice, because the data sits next to your rankings rather than in a separate login.
• Crawler and log analytics. Identify and verify AI bot traffic in server logs — the access question rather than the citation question.
• Referral attribution. Analytics configured to isolate visits originating from AI products, usually via referrer patterns.
• Build-it-yourself. Scripted prompts against provider APIs, logged to a database. Cheapest at scale, with the caveat that an API response is not always identical to what a person sees in the consumer product.

Two things to check before signing anything. Ask how the vendor obtains its answers — API calls, headless browsers, or a panel — because that determines how closely the data matches real user experience. And ask whether you can supply your own prompt list; a tool that only reports on prompts it chose is measuring its own idea of your market.

When the engine gets you wrong

Most measurement frameworks treat this as a curiosity. It is frequently the most urgent finding.

Ask the engines directly what your company does, what it charges, where it operates, and whether it's any good. You will sometimes find a service you discontinued described as current, pricing that was never yours, a location you don't serve, or confident confusion with a similarly named company — delivered in the same authoritative tone as everything else, to a reader with no way to know it's wrong.

That's worse than absence. Being unmentioned costs you a chance. Being misdescribed costs you buyers who now believe something false and have no reason to check.

Treat every inaccuracy as a work item with an owner and a due date, and fix it at the source rather than complaining to the model:

1. Find the origin. Misinformation usually traces to something real — a stale page on your own site, an outdated directory listing, an old press release, a third-party profile nobody has updated, or a name collision with another company.
2. Correct your own site first. Update the page, and make the correct facts unambiguous in your structured data so the entity resolves correctly rather than by inference.
3. Correct the third parties. Directory listings, industry databases, review platforms, your own social profiles. Engines synthesize across sources; a single stale profile can keep resurfacing.
4. Publish the correct answer explicitly. If an engine keeps saying you serve a region you don't, publish a page that states plainly where you do and don't operate. Retrieval needs something to retrieve.
5. Re-check on your monthly cycle. Corrections take time to propagate as engines re-crawl and re-index. Verify rather than assume.

Every one of these is slow, and none of them is a guarantee. Nobody controls what a model says. What you control is the quality and consistency of the evidence available to it — which is the same thing that makes you citable in the first place, described in answer engine optimization.

Frequently asked questions

What is AI visibility?

AI visibility is the extent to which AI answer engines — ChatGPT, Perplexity, Claude, Google's AI Overviews and AI Mode — name your business when someone asks a question you should be the answer to. It replaces ranking position as the relevant measure, because these products return one synthesized answer with a few named sources rather than a page of ten links. It is measured in mentions and citations rather than positions, and because the answers are non-deterministic, it is measured as a rate across a fixed set of prompts over time rather than as a single observation.

What's the difference between a brand mention and a citation?

A mention is your brand named in the text of an answer. A citation is your domain appearing as a linked source the answer was built from. They have different causes and different values. Mentions can come from the model's general knowledge of your company, so they reflect how widely you're written about across the web. Citations require your specific page to have been retrieved for that query, so they reflect crawlability, page structure and topical fit. Only citations can send you traffic, but mentions still shape what a buyer believes before they ever click anything.

How often should I check AI visibility?

Monthly, on the same week, with the same prompt set and the same session conditions. Weekly checking mostly measures the non-determinism of the systems rather than any change in your standing, and it burns effort producing charts that jitter. Quarterly is too slow to catch an accuracy problem or a crawler block before it does damage. The trend only becomes readable after three or four runs, which is the point at which most programs are abandoned. Hold the prompts constant even when they start to feel stale — changing them resets your baseline.

Can I get an exact ranking position in ChatGPT?

No, and any tool reporting one is presenting an estimate as a fact. AI answers do not have ordinal positions. The same prompt can produce different sources and different companies named across sessions, users and model versions, and the engines do not publish retrieval logic or expose a ranking. What can be measured honestly is frequency: across a fixed prompt set run repeatedly, how often you appear, how often you're cited with a link, and what share of total brand mentions you hold against a defined competitor set. Treat any single-number "AI rank" as a marketing construct.

What should I do if an AI engine describes my business incorrectly?

Trace it to a source rather than trying to correct the model. Inaccuracies almost always originate in real content — a stale page on your site, an outdated directory or review profile, an old press release, or confusion with a similarly named company. Fix your own pages first and make the correct facts explicit in your structured data so your entity resolves unambiguously. Then update the third-party listings, since engines synthesize across sources and one neglected profile can keep resurfacing. Publish a page that states the correct answer plainly, then re-verify on your normal monthly cycle. Propagation takes weeks, not days.