How to track your visibility in AI search tools

With ordinary SEO, you can always answer the basic question: where do we rank? Type the query, look at the page, done. With AI search there's no page to look at. An assistant composes each answer fresh, names a few businesses, and the moment passes without leaving a public record. Most owners conclude the whole thing is unmeasurable and stop thinking about it, which is exactly wrong: the visibility is measurable, just through different instruments, and businesses that measure it early get months of head start on an increasingly consequential channel.

The trick is knowing what's observable. AI search visibility shows up in three layers: your analytics, your server's visitor log, and the answers themselves. Two of the three are free.

First, understand what you're measuring

One calibration before the tools. AI answers are not stable the way rankings are. The same question, asked twice, can produce different brand lists, and phrasing changes the answer again. So you're not tracking a position, you're sampling a distribution: how often, across repeated realistic questions, does your business get mentioned or cited? Think polling, not scoreboard. That mindset determines the method: repeated checks, consistent questions, trends over months. One-off spot checks tell you almost nothing except when the answer is badly wrong about you, which is worth knowing immediately.

Layer 1: the traffic that reaches your analytics

The most direct evidence is people arriving at your site from AI tools. When someone clicks a citation in ChatGPT or Perplexity, they arrive as a referral from hostnames like chatgpt.com or perplexity.ai, and GA4 records it. GA4 now even sorts major assistants into their own "AI Assistant" default channel group. If yours doesn't show it yet, an exploration filtered on session source containing chatgpt, perplexity, copilot, gemini, and claude gives you the same view in five minutes.

Look at three things: the trend (AI referrals are small for almost everyone, but the slope matters), the landing pages (which tells you which of your content AI systems consider citable), and what those visitors do, since conversion, not volume, is the point.

Now the crucial caveat: this layer badly undercounts your real AI visibility. Pew's research found that when Google shows an AI summary, only about 1% of visits produce a click on a source inside it, and Similarweb clickstream research, reported by SparkToro, found that people who get an AI recommendation often don't click anything at the time, they later search your name or type your address directly. Which means two of your oldest metrics have become AI-visibility instruments: branded search volume in Search Console and direct traffic in GA4. A rise in either, without a campaign to explain it, is frequently the shadow of recommendations you never saw.

Layer 2: the crawlers in your server log

Before an AI system can cite you, it has to read you, and the reading is observable. The major AI companies operate crawlers with published identities. OpenAI documents its bots with exact user-agent strings and public IP lists, and, importantly, splits them by job: GPTBot gathers training data, while OAI-SearchBot fetches pages to build ChatGPT's search answers. Anthropic and Perplexity publish their crawler identities similarly.

Two checks follow. First, your robots.txt: make sure you aren't blocking the search-answer bots, possibly by accident via an overzealous "block all AI" rule added when the concern was training data. Blocking training crawlers is a legitimate choice, but blocking the retrieval bots takes you out of the citation game entirely, and the two are controlled separately. Second, your server logs or CDN dashboard: are these bots actually fetching your pages, and which ones? AI crawlers visiting your key pages is the upstream signal that the assistants consider you source material. Silence is a finding too, usually pointing at technical accessibility or a robots.txt problem.

Layer 3: the answers themselves

The most informative layer is also the most manual: ask the assistants what they say about your market, and write down the results.

Build a fixed panel of ten to twenty questions that mirror how your customers actually ask: "who's a good [your service] in [your city]?", "best [your product] for [your customer type]", "[competitor] vs alternatives", "is [your business name] reputable?". Run the panel monthly on the platforms that matter, ChatGPT, Perplexity, Google's AI Overviews and AI Mode, and, because answers vary run to run, ask each question fresh rather than reading one long conversation. For each answer, log three things in a spreadsheet: were you mentioned, were you cited as a source, and was what it said accurate?

Thirty minutes a month yields the dataset nobody else in your market has: your share of the answers, trending over time, plus every factual error an assistant tells customers about you. Errors are gold, because they're fixable: wrong facts trace back to stale or inconsistent sources about your business, and correcting the source corrects future answers.

For the Google-specific slice, Search Console holds a hidden version of this data. AI Overview impressions and clicks are folded into the ordinary Performance report, with no separate filter, so you can't see them labeled, but you can see their signature: question-type queries where impressions hold steady or rise while clicks sag. Those are the queries where an AI answer now sits between you and the click, and they're your priority list for earning the citation instead.

The panel spreadsheet, concretely

Since the question panel is the heart of the system, here's the exact artifact, buildable in fifteen minutes.

Columns: the question, the platform, the date, three yes/no columns (mentioned, cited as a source, facts accurate), a "who was recommended instead" cell, and a notes cell. Rows: your ten to twenty questions times the platforms you check. Write the questions the way customers talk, not the way marketers do: "is it worth getting a heat pump serviced every year", not "heat pump servicing benefits". Include at least two questions about your business by name, two naming your main competitors, and a spread across the buying journey from "how do I fix X" to "who should I hire for X".

The derived numbers you'll actually watch: mention rate (the share of answers naming you), citation rate (the share linking or crediting your pages), and accuracy rate for the answers that discuss you. Three percentages, tracked monthly, graphed in the same spreadsheet. The competitor column earns its place the first time you notice one rival appearing in eight answers out of ten and can go study which sources every one of those answers cited.

Two habits keep the data honest. Ask fresh, one question per new conversation, since assistants adapt within a chat and a long session stops resembling what a new customer sees. And don't chase single-month wiggles: these systems are stochastic, meaning the same question legitimately produces different answers on different runs, so judge movement on two or three months of direction, the same discipline any small-numbers metric demands. The spreadsheet's job isn't precision, it's making a previously invisible channel visible enough to manage.

The paid layer, and when it's worth it

What the manual panel does at spreadsheet scale, commercial tools do industrially: Semrush's AI toolkit and Ahrefs' Brand Radar sample enormous numbers of prompts and track brand mentions, citations, and share-of-voice across platforms over time, with competitor comparisons built in.

The honest small-business guidance: start manual. The spreadsheet version costs nothing, teaches you what the tools would be measuring, and for a local or niche business, twenty well-chosen questions cover most of the ground that matters. Graduate to a paid tracker when one of three things becomes true: you're competing in a broad national category where twenty prompts can't represent the question space, you need competitor share-of-voice to make a strategic case, or AI referrals have grown into a channel whose budget decisions need defending with data. The tools are real; they're just rarely the first dollar to spend, and budget SEO principles apply unchanged.

The monthly routine, assembled

Pulling the layers into one recurring half hour:

• Check the AI referral segment in GA4: trend, landing pages, conversions. Five minutes.
• Glance at branded search impressions in Search Console and direct traffic for unexplained rises. Five minutes.
• Run the question panel and update the log: mention rate, citation rate, accuracy. Fifteen minutes.
• Quarterly, add ten minutes: confirm the AI search bots are fetching you (logs or CDN), and re-check robots.txt after any site change.

Then act on what the measurements say, because each layer points at its own fix. No crawler visits: fix access. Crawled but never cited: your content isn't the kind that gets quoted, usually meaning it lacks extractable answers, data, or authority. Cited but factually wrong: fix your inconsistent sources. Mentioned but losing to competitors: study what sources the assistants cite for them, and earn presence in those same places.

Quick answers to the follow-ups

Which platform should I care about most? Weight them by your customers, not by tech news: for local services, Google's AI surfaces matter most because that's where local buyers already are; for younger or technical audiences, ChatGPT and Perplexity deserve fuller panels. Your GA4 referral data settles the argument over time with your own numbers.

My business never appears anywhere. Is something broken? Usually nothing is "broken", the assistants just have thin evidence about you. Work the layers in order: confirm crawler access, then fix your entity consistency and business profiles, then publish the extractable, specific pages that give the machines something to quote. Absence is the starting condition for most small businesses, not a penalty.

Can I just ask ChatGPT why it didn't recommend me? You can, and it will produce a plausible-sounding explanation, but treat it as brainstorming, not diagnostics: models don't have reliable access to their own selection reasons. The citation studies and your own source-tracing of competitors' mentions are the real diagnostic.

Is any of this worth it at my size? The free monthly routine costs half an hour and protects against the one scenario nobody can afford: assistants telling your market something wrong about you while you never check.

The one-line version

You can't see a ranking in AI search, but you can see everything that matters: who arrives from the assistants, whether their crawlers read you, and what the answers actually say when asked your customers' questions. Measure those three layers monthly, free, in half an hour, and you'll know your AI visibility better than competitors twice your size, while it's still early enough for that knowledge to be an advantage.