Iron Goo
P3E04 · Part 3: AI SEO

AI visibility: track appearance rates, not rankings

Measure AI visibility without mistaking citations for traffic or a single answer for a trend. Learn how to combine free reports, verified server logs and repeated prompt tests into a weekly SEO tracker.

What this episode covers

  • Google and Bing AI reports, what they measure and where their limits matter.
  • Grounding queries that reveal content gaps and help you improve existing pages.
  • Verified AI page fetches and the difference between being read and earning a visit.
  • Referral tracking by host so internal clicks do not inflate your AI traffic.
  • Fixed prompts in customer language, repeated tests and appearance rates measured over weeks.
  • Separate columns for brand mentions and source citations, with dates and trends you can report clearly.
  • Why useful tools can earn clicks when complete answers stay inside the chat.

Key moments

  1. 00:00Intro
  2. 00:57The free first-party data
  3. 03:35Logs and referrals
  4. 06:10Prompt tracking, done right
  5. 08:32Who writes the prompts controls the score
  6. 09:28Mentions are not citations
  7. 10:31Two verdicts
  8. 11:29Your weekly tracker and wrap up

Ready to move?

Tell us where your business is today. A written assessment lands in your inbox within two business days.

Talk to us

Transcript

Intro

Microsoft's AI cited an education site we ran seventeen thousand seven hundred and forty eight times. Over the site's entire analytics history, Copilot sent it about twenty six visitors. Both numbers are real, and both came from free tools. So which one is your AI visibility?

Depends what you want to know. Citations tell you the machines trust your pages. Visits tell you whether people came. Most AI visibility dashboards blend the two into one score, and that score can be pushed up or down by whoever writes the prompts behind it. Today we build a setup that can't flatter you.

Four layers, all cheap or free: Google's and Bing's own AI reports, your server logs, referral visits counted correctly, and prompt tracking done honestly. Then two verdicts on what gets sold as AI rank tracking.

The free first-party data

Start with Google. Search Console's generative AI performance reports reached every site on the thirty first of August 2026. They show how often your pages appeared in AI Overviews, AI Mode and Discover's AI features, by page, country, device and date.

Read the limits before you report from it. Impressions only, no clicks, no queries. And Google says these impressions already sit inside your normal performance totals, so adding the two together counts them twice. Yes, someone will do that in a board deck this quarter.

Used well, it still answers a useful question: which of my pages show up in Google's AI features, and is that growing? Put its page list beside your normal clicks for the same pages. A page with rising AI impressions and flat or falling clicks is a page whose answer is being read somewhere else.

Bing gives you more, and almost nobody looks. Since February 2026, Bing Webmaster Tools has an AI Performance report covering Copilot, Bing's AI summaries and some partners. It shows citations per page, plus grounding queries: the search phrases the AI used to go and fetch your content. In June, Microsoft added intents, topics and citation share.

Here is what it showed for that education site. Lessons took ninety percent of the citations. The listed grounding queries covered only forty two percent of the total, because Bing leaves out sparse queries. And on those queries, the site held about a fifth of all citations shown. That last number is the closest thing to a share of voice you'll get for free.

The grounding queries are the part to steal for your content plan. They are the actual searches an AI ran before it quoted you, in the AI's own words. If a grounding query lands on a page that only half answers it, that's a section to write. If it lands on the wrong page entirely, that's an internal link or a clearer heading to fix.

Microsoft is refreshingly blunt about what these counts are not. They don't tell you ranking, authority or where you sit inside an answer, and Microsoft says the new views don't turn AI visibility into a single score. Good. A citation count is a direction to watch, not a league table.

One practical note for next episode's workbench. As of October 2026, the Search Console API has no generative AI search type. Web, image, video, news, Google News and Discover, yes. AI report, no. So for now you export it from the interface by hand, the old-fashioned way.

Logs and referrals

Layer two is your own server logs. When someone asks Claude, ChatGPT or Perplexity a question and the assistant opens your page to answer it, the fetch arrives with a user agent like Claude-User or ChatGPT-User. Each one is a sign that a real person's question touched your page. Verify the IP address against the vendor's published list first, because fake bots are everywhere. Next episode shows how many.

Two setup traps before you trust a single log line. If your site sits behind a CDN, your logs may hold the CDN's addresses instead of the visitor's, and then no bot can be verified. Restore the real client IP first. And store counts per page and per day, never IP addresses or anything about a person.

Now put fetches next to visits and you get the ratio that matters: AI referral visits per hundred AI fetches. On an education site we run, over two weeks in September 2026, topic guides scored zero point eight. An interactive test scored twenty three point five. Four visits on each side, and the direction is clear.

The direction is still useful. Pages that answer fully inside the chat get read and not visited. Pages that offer something a chat can't do, like a test, a calculator or a live listing, get the click. Episode one showed the same pattern in Google's own AI summaries, from the user's side.

Layer three is referral visits, the people who actually click through. OpenAI says ChatGPT adds utm_source equals chatgpt.com to the links it sends, which makes them easy to spot. It also makes them easy to miscount. The classic mistake is counting any visit whose referrer mentions an AI site anywhere in the URL.

Here's why. A visitor lands from ChatGPT, and the tagged URL stays in their address bar. Every click they make inside your site now carries it in the referrer. A widely shared AI traffic snippet falls for exactly this. When we ran it on an education site's logs, eight hundred and thirty five of its eight hundred and forty so-called AI referrals were internal clicks.

We checked again last week on the same site, from the twelfth to the eighteenth of October. One hundred and sixty two log lines had an AI domain somewhere in the referrer. All one hundred and sixty two came from the site's own pages. The fix is one rule: match the referrer's host, never the full URL.

Prompt tracking, done right

Now the layer everyone sells: prompt tracking. You ask an assistant a question and record whether your brand shows up. Sounds simple. So we tested how stable the answer is, by asking Claude the same question twenty times on the nineteenth of October 2026: what are the best websites to check where a movie is streaming in the US?

JustWatch and Reelgood appeared in all twenty answers. Google search was recommended in fifteen, Yidio in ten, Plex in seven, IMDb in three. Six more names turned up once or twice. Twenty runs, thirteen different lists.

Now imagine you're Yidio and your tool ran this prompt once. Half the time you'd report that you're winning, and half the time that you've vanished, from the same question on the same day. Same brand, same afternoon, two completely different board meetings.

The large studies agree. One study ran the same brand questions almost three thousand times late in 2025, and the chance of getting the same list of brands twice was under one in a hundred. Another study asked ChatGPT the same seventy three questions every day for two months. On most days, at least half of the previous day's mentions were gone.

There is good news in that data too. The same five names held the top five in both halves of the two months. The top of the list is stable, and the tail churns, which is exactly what our twenty runs showed. So measure appearance rate over many runs and many weeks, and expect the tail to wobble.

One rule makes or breaks it: freeze the prompt set. The moment someone edits a prompt because the brand didn't show up, the trend line is measuring the edit, not the AI. If a prompt is genuinely bad, retire it, add a new one, and start that line from zero.

Appearance rate is simple arithmetic: runs where you were named, divided by all runs. That day, Yidio's rate on our question was fifty percent. Run the same prompts every week, track your rate and two or three competitors', and judge the trend over a month or more. One week tells you very little.

Running this doesn't need an enterprise tool. Our twenty runs went through the API from Claude Code with web search switched on, and the whole batch cost under a dollar. One setup note: the API isn't identical to the consumer app, so state which one you measured, and keep it the same every week.

Who writes the prompts controls the score

Here is the part the dashboards leave out. One SEO agency tested its own brand in July 2026. When the prompts used the category words an SEO agency would use, it was named about twenty four percent of the time. When the prompts described a buyer's situation without the industry label, it was named under four percent of the time.

That's a six and a half times swing, mostly from wording, and the authors are careful to say the prompt groups also differed in length and detail. The lesson stands either way. Whoever writes the prompt set controls the score. A tool that won't show you its prompts is showing you a number it chose.

So write your prompts from customer language, not yours. Pull the wording from sales calls, support tickets, reviews and forum threads, and include questions that never mention your category at all. For ReelRadar, that means 'what can I watch tonight on the services I already pay for', not 'best where to watch aggregator'.

Mentions are not citations

One more trap: being named and being cited are separate events. Semrush and a partner analysed nearly four thousand brand appearances in AI answers in June 2026. In about six out of ten, the brand's site was cited as a source and the brand was never named. ChatGPT cited a brand's site most of the time it appeared, and named the brand only about a fifth of the time.

Our twenty runs showed the mirror image. JustWatch was recommended in all twenty answers, and justwatch.com was cited as a source in none of them. Meanwhile a small blog about where to watch films was cited in fourteen runs, and the app it promotes was recommended in only two. Track mentions and citations as two separate columns.

What do you do with citations that don't name you? Make your name hard to leave out. Use one brand name and one short description of what you do, the same everywhere, on your own pages and on every profile. An assistant quoting your page has a better shot at naming you when the page says plainly who you are.

Two verdicts

Claim one. AI rank tracking tells you where you rank in ChatGPT. Plenty of tools sell a position number, just like a rank tracker for Google, often with a nice green arrow next to it.

There is no stable position to track. The same prompt gives a different list almost every time, the order moves even more than the list, and wording alone swings appearance six fold. Appearance rate over many runs is measurable. A single rank is noise with a decimal point.

Claim two. One screenshot proves we're winning, or losing, in AI search. You've seen these in pitch decks, and in panicked late-night emails from the boss.

A screenshot is one run of a system that changes its answer half the time. In our test, Yidio could have produced a winning screenshot or a losing one from the same question, on the same day, ten times each. Ask for the prompt, the number of runs and the date range, or ignore the screenshot.

Your weekly tracker and wrap up

Here's the weekly tracker we use, free on the episode page. One tab per source: Search Console AI impressions by page, Bing citations and grounding queries, verified user fetches from your logs, and AI referral visits matched by host. The last tab is your prompt set, run several times a week and logged as appearance rate, with mentions and citations in separate columns.

Then read the layers against each other, page by page. Lots of citations and few visits means the answer is being absorbed, which is fine for brand exposure and bad for traffic. If you want the click, give that page something a chat can't do for the reader.

When you report it to a client or a boss, keep it to one page with dates on every line: AI impressions, Bing citations, verified fetches, AI referral visits, and your prompt set's appearance rate. Show the trend and the period. Never a single blended score.

The episode on one screen. Use the free reports from Google and Bing, verify your log fetches, count referrals by host, and track prompts as appearance rate over weeks. AI rank tracking and the single screenshot both came out as placebo.

What will change: the reports Google and Bing offer, and how often assistants reshuffle their answers. What won't: fixed prompts in customer language, many runs, separate columns for mentions and citations. The prompt set template and weekly tracker are free at irongoo.com. Next episode, we build the workbench that pulls all of this for you.

Free ebook

Put your next steps in order

Build on what you have just read with a free ebook on semantic SEO and AI visibility. Get worked examples, decision rules, checklists and a do-this-week action in every chapter, plus a six-week plan for your site.

By signing up and confirming your address, you consent to receive the free ebook plus one email every two weeks with 3 or 4 new Iron Goo videos, guides and articles, and nothing else; unsubscribe any time with one click.