How do I track whether AI tools mention my business?

Start by accepting an uncomfortable fact: nobody can tell you exactly how often AI tools mention your business. Google gives you Search Console — every query, every impression, every click. OpenAI, Anthropic, Google's Gemini team, and Perplexity give you nothing equivalent. There is no dashboard, no API, no impression count. The conversations where you're recommended (or your competitor is) happen privately, between a user and a model, and leave no trace you can access.

It gets harder. AI answers are non-deterministic: ask ChatGPT 'best plumber in North Vancouver' twice and you can get two different lists. The answer varies by user, by session, by memory and chat history, by phrasing, and by whether the model decided to run a live web search that time. So even your own spot-checks only sample a distribution — they don't reveal a fixed 'ranking' the way position 3 on Google does.

And the biggest blind spot: most AI exposure produces zero referrer. When ChatGPT tells someone 'SearchPod Inc. is a Toronto agency that manages Google Ads for SMBs' and the user simply acts on that — calls, or Googles the name — your analytics never see the AI's role. The mention happened; the measurement didn't. Every method below works around these three problems rather than solving them. Combined, they give you a usable picture. Just don't expect Search Console-grade precision, because it doesn't exist yet for anyone.

When a user clicks a link inside an AI answer, the visit lands in GA4 with the assistant's domain as the referrer. Build a custom channel group (or an exploration with a referrer filter) that matches the major AI domains: chatgpt.com and chat.openai.com (ChatGPT), perplexity.ai (Perplexity), gemini.google.com (Gemini), claude.ai (Claude), and copilot.microsoft.com (Copilot). Name it something like 'AI Referral' and it becomes a trackable channel alongside Organic Search and Paid.

This is the only fully automated, zero-effort signal on this list, and it's worth fifteen minutes to set up today — even if the numbers start tiny. For most SMBs they will start tiny, often single digits per month. That's normal and not a verdict on your AI visibility.

The caveat to internalize: referral traffic undercounts massively. AI assistants are answer engines — the whole point is that the user often doesn't need to click anything. People read the recommendation, then phone you, walk in, or search your brand name on Google (which lands in Organic or Direct, not AI Referral). Treat GA4's AI channel as a directional floor — useful for trend, useless as a total. If it's growing quarter over quarter, your real AI exposure is almost certainly growing faster.

Since you can't see other people's AI conversations, simulate them. Here is a simple method.

Step 1: Write a fixed list of unbranded prompts. Unbranded means your business name is not in the prompt, because that is how new customers ask: 'best [your category] in [your city]', 'who should I hire for [service] in [city]', 'how much does [service] cost in [city]'. Keep branded checks like '[your business] reviews' on a separate list.

Step 2: Pick a set schedule, such as the first Monday of each month. Run every prompt in a fresh session (logged out or a new chat with memory off).

Step 3: For every run, write down: the date; the engine and the model (for example ChatGPT and the model name it shows); whether web search was on; whether your business was mentioned; and which URL was cited, if any (yours or someone else's). Add a notes column for anything wrong, like an old address.

Step 4: Keep leads separate. In another tab, log calls, forms and bookings where the person said they found you through an AI tool, plus AI referral visits from GA4. Do not mix leads into the mention log. A mention is not a lead.

In ChatGPT, answers that use web search show their sources, so you can see which URLs were cited. OpenAI explains this in Searching the web with ChatGPT. When web search is off, there may be no citations at all, which is why that column matters.

Discipline matters more than tooling here. Use logged-out or fresh sessions where possible so your own chat history doesn't contaminate results — assistants personalize, and a model that knows you work at the business will flatter you. Keep the prompt wording identical month to month; changing the questions destroys the trend line. The first month is your baseline and will probably sting. That's fine — the panel exists to show movement, and movement is what you're buying with all your GEO work.

Citations are the closest thing AI search has to rankings. When Perplexity or ChatGPT's search mode answers a question, it lists the sources it drew from. Tracking which of your URLs appear as cited sources — and which third-party pages about you get cited — tells you what the engines actually trust. During each run, log every cited URL, yours or not. For a quick sample in one engine, SearchPod's free Perplexity citation checker asks Perplexity up to 5 questions one time and shows whether your domain was cited. It does not check ChatGPT, and results change by day and session, so treat it as a spot check, not a score.

The citation list doubles as your to-do list, which we'll come back to below: every source the engines cite is a place you either need to be, or need to beat.

Crawler logs show access, and only access. Bots like OAI-SearchBot and GPTBot (OpenAI), ClaudeBot (Anthropic) and PerplexityBot name themselves in the user-agent string, and OpenAI lists its bots and what each one does at platform.openai.com/docs/bots. Pull your server or CDN logs and search for them. A hit means a bot fetched a page. It does not mean the page was indexed, cited, or recommended. And no hits does not prove you are left out: some fetches never reach your logs (for example, cached copies or third-party sources), and an engine can name you from what other sites say. What crawler logs are good for is spotting blocks. If you see no hits at all, check robots.txt, your firewall and any bot-protection layer. It's the cheapest health check on this list.

Remember the zero-referrer problem: most people who hear about you from an AI don't click a link — they Google your name. So branded search volume and direct traffic become your lagging indicators of AI exposure. Watch branded impressions and clicks in Search Console, and direct-traffic trends in GA4. If your prompt-panel results are improving and, a month or two later, branded search climbs without any other explanation (no PR, no campaign, no seasonality), the 'asked ChatGPT, then Googled you' pipeline is the likely cause. It's circumstantial evidence, but over quarters it's persuasive — and it's the signal that actually correlates with revenue.

Now, the tools question, answered honestly. A category of dedicated AI-visibility trackers has emerged — platforms that run prompt batteries across assistants automatically, track your share of voice against competitors, and monitor citations at scale. The category is young: methodologies differ, coverage of assistants varies, pricing is all over the map, and because AI answers are non-deterministic, two tools can report different numbers for the same brand in the same month. None of them has access to real user conversations either — they're automating the same sampling you'd do by hand, just at higher volume.

For an enterprise tracking hundreds of prompts across markets, they can earn their fee. For a typical SMB, a spreadsheet plus monthly discipline beats most paid tools: it's free, you understand exactly what was measured, and 10–20 prompts is genuinely enough to see your trend. Graduate to paid tooling when the manual panel becomes the bottleneck, not before.

Measurement only matters if it changes what you do next, and the highest-leverage output of all this tracking is the cited-sources roadmap. Every URL that AI answers cite in your category is a vote about what the engines trust. If Perplexity keeps citing a directory you're not listed in, get listed. If ChatGPT builds its answer from a competitor's pricing page, publish a clearer one. If a 'best of' roundup is cited across multiple assistants and you're absent from it, pitching that one publication may move your AI visibility more than months of on-site work. You're reverse-engineering the engines' trusted-source list and systematically joining it.

The panel also catches accuracy problems. If a model states your old address, a discontinued service, or a wrong price, that's usually traceable to a stale page somewhere on the web — your own site, an unclaimed profile, an outdated directory entry. Fix the source and the answers follow, eventually.

On cadence: monthly is enough. AI visibility moves on the timescale of crawls, index refreshes, and model updates — weekly checking adds noise and anxiety, not insight. Expect high variance between runs: being named in three of four assistants one month and two the next is normal sampling noise, not a crisis. Judge the three-month trend, never a single answer. And set expectations accordingly — a quarter of flat results while you build citations and clean up your entity information is common, then improvements tend to arrive in steps rather than a smooth curve. The discipline of measuring the same way, every month, is what turns an unmeasurable channel into a manageable one.