How to Measure AI Visibility by Topic Now That Bing Reports It That Way
My answer to how to measure AI visibility right now is to group your prompts into topics, track how often your brand shows up across each whole group over a couple of months, and put very little weight on where you land for any single prompt on any single day. Microsoft started reporting AI citations by topic inside Bing Webmaster Tools this month, and the bigger studies from this year point in the same direction.
The Bing piece is the news. On October 13 the Bing Webmaster Tools team added a beta feature called Topics Portfolio to its AI Performance report. It takes the grounding queries your site gets cited for and clusters them into themes, so you can see, in Microsoft's words, "the shape of your presence across those clusters rather than one query at a time." As far as I know it's the first time one of the engines has handed site owners their own AI citation data organized by topic.
The short version
Single prompts make a bad unit of measurement, because AI answers change from run to run and one check tells you very little.
Topics, measured as a share of answers over several months, give you a number that holds still long enough to manage.
Most categories in ChatGPT don't have a clear leader yet (53.7% were unsettled in a Semrush and Growth Memo study), so there's still room to move.
Bing's Topics Portfolio is free first-party data, and scraped tracking data deserves a little more suspicion than it used to get.
What Bing's Topics Portfolio shows you
If you already use the AI Performance report, which covers citations in Microsoft Copilot, Bing's AI summaries and some partner integrations (Search Engine Land), you'll find Topics Portfolio as a new tab next to Citations & Pages. It only appears once your site has enough data.
Each topic gets a handful of numbers. Topic Size is the approximate total citations across all sites for that topic over the period you pick (one, two or three months). Sites Cited is roughly how many sites are showing up for queries in that topic. Queries is the count of distinct grounding queries in the topic. Then there's a list Bing calls Top Opportunity Queries, which are strong queries where other sites get cited and yours doesn't (Bing Webmaster Blog).
That last list is the part I'd spend time on. It's Microsoft telling you, from its own logs, where the engine picked somebody else in a topic you already play in. I covered the AI Citation Share metric when it landed in Bing Webmaster Tools back in July (it's in my post on how to read an AI visibility study), and this is a step further, because it gives you the gaps as well as the score.
Microsoft is careful about how much to trust it. The topic grouping is done by AI and "can occasionally be imperfect," and the post says to "treat the numbers as directional signals for planning, not as a precise ledger." It's also Bing and Copilot only, so it won't tell you anything about ChatGPT or Gemini. Still, it's first-party and free, and it's organized around topics.
Why one prompt at a time gives you bad numbers
If you've ever been asked "are we showing up in ChatGPT?" you know the temptation. You type the question in, screenshot the answer and put it in a deck, and the same question asked again a few minutes later can come back with a different list.
Earlier this year Rand Fishkin at SparkToro and Patrick O'Donnell at Gumshoe had about 600 volunteers run 12 identical prompts through ChatGPT, Claude and Google's AI nearly 3,000 times. The chance of getting the same list of brands twice was under 1 in 100, and the same list in the same order was closer to 1 in 1,000. Some brands still showed up in 60% to 90% of responses for a given intent, even while their positions moved around (Search Engine Land). Their advice was to track how often your brand appears across many prompts, run many times.
My own AI Verdict Study found the same kind of wobble further down the funnel. I held the buyer, the brand and a seven-turn script identical across runs, and 21 of 72 brand-engine pairings (29%) still flipped between recommending a brand and rejecting it.
So a single check is a sample size of one from a system that's designed to vary, and you can't report a month-over-month change on that to a leadership team with a straight face.
Three ways to count AI visibility
Pick a view to see what each method measures.
"What's the best payroll software for restaurants?"
Example lists for illustration.
Chance of getting the same brand list twice from the same prompt, across ChatGPT, Claude and Google's AI. Same list in the same order was closer to 1 in 1,000.
One topic = five prompts that cover how people ask
× several runs each × every engine your buyers use
Some brands showed up in 60% to 90% of responses for an intent even while their positions moved around. Presence across many runs is the number that holds still.
What each topic shows in Bing Webmaster Tools > AI Performance
Covers Copilot and Bing AI answers only. Bing calls the numbers "directional signals for planning."
Sources: SparkToro and Gumshoe via Search Engine Land; Semrush and Growth Memo (five-prompt topic design); Bing Webmaster Blog, Oct. 13, 2026.
How to measure AI visibility by topic
The Semrush study I mentioned up top, done with Kevin Indig and Growth Memo, is a good public model for doing this by topic. They tracked 1,094 US categories in ChatGPT every month from January through June 2026, covering more than 50,000 brands. Each category was a cluster of five prompts written to cover the way people ask: a definition question, a comparison, an alternatives question, a use case, and a buying question (Semrush).
They called a brand the category owner if it showed up in at least four of the five prompts with a lead of five points or more over the runner-up. By that bar, 15.2% of categories had an owner, 31.2% had an emerging leader, and 53.7% were unsettled, meaning no brand made it into even three of the five prompts. The busier topics were less settled, too. Only 11.3% of the top half of topics by AI search demand had a clear owner, against 19% of the bottom half.
A few other things from that study change how I'd set up a report. Clear owners stayed on top in 90.4% of month-over-month comparisons, so once a brand opens a wide lead in a topic it tends to keep it. In the emerging and unsettled categories the top brand changed in 1,950 of 5,470 comparisons, and the leader flipped most often when the gap was tiny (a median lead of 1.3 points, against 2.9 where the leader held). Standard SEO numbers didn't predict much either. Owners had more organic traffic than the runner-up in only 48.4% of pairs, which is about a coin flip.
Then there's the gap between being mentioned and being cited. Only 21% of the most-cited domains in a category were also the most-mentioned brand. I wrote a whole post on why ChatGPT can recommend you and still not cite your site, and this is the topic-level version of the same split. Bing's report counts citations, so you'll need your own testing for mentions.
Most ChatGPT categories have no clear leader
1,094 US categories, five prompts each, tracked monthly January to June 2026.
Source: Semrush with Kevin Indig and Growth Memo, "AI visibility is a topic-level game," July 2026. Owner = top brand in at least 4 of 5 prompts with a 5+ point lead. Unsettled = no brand in 3 or more prompts. "Leader held" for other categories calculated from 1,950 changes in 5,470 month-over-month comparisons.
Scraped tracking data has its own problem right now
There's one more reason I'd lean on first-party data like Bing's where you can get it. On October 14 Search Engine Journal wrote up an analysis from the rank-tracking company Nozzle showing that since September 20, Google has been serving some automated rank-checking requests what Nozzle calls poisoned results. At the peak it hit 25% of results, and it's been running at 5% to 11% a day since October 1. Normal results kept 88.5% of the same sites when the same keyword was searched again, and the poisoned ones kept 20.7% (Search Engine Journal).
AI Overviews showed up in 34.9% of the poisoned pages against 54.2% of normal ones, so if your AI Overview tracking runs off scraped results, some of it may be undercounting. I don't know how far this goes beyond Nozzle's data, and the article includes no comment from Google. I'd pair any third-party tool with whatever the engines will tell you directly, and Google's version of that is still taking shape, which I got into in the post on Google's AI contribution pilot.
A topic scorecard you can build this month
If I were starting from scratch, this is how I'd set it up.
Number one, pick somewhere between five and ten topics that match how your buyers shop. If you sell payroll software, "payroll for restaurants" and "switching payroll providers" are topics, and your product line names probably aren't.
Two, write five prompts for each topic using the Semrush shape: what is it, how does A compare to B, what are the alternatives to the leader, which one is good for this situation, and which one should I buy.
Three, run every prompt more than once on every engine your buyers use (ChatGPT, Gemini, Perplexity, Google AI Mode and Claude, for most of us), and record whether you were mentioned and whether you were cited. Skip position entirely. Your number for each topic is the share of answers you appeared in.
Four, report the topic number monthly and look at it over at least two or three months before you call anything a trend. Narrow leads in busy topics flip from month to month in the Semrush data.
Then last, open Bing's Topics Portfolio once a month, line its topics up against yours, and use the Top Opportunity Queries list as a content backlog. Go read what's getting cited there and figure out what it has that your page doesn't.
If you'd rather not build the sheet yourself, the AI Visibility Audit Kit includes a Buyer Prompt Pack with 40 prompts across 10 categories, a run log, and a scorecard that compares your first round of results with your latest, by engine. If you want a quicker gut check first, the free Quick Check takes about 20 minutes. My post on AI search visibility in 2026 has the bigger picture on why being in the answer counts for more than the click now, and Where AI Engines Go to Decide About Your Brand covers the sources the engines lean on.
Does your AI visibility report measure topics?
Seven questions. Answer for the report your team uses today.
Answer all seven to see your score.
What I'd like to hear from you
The topic approach has an open question I can't answer yet, which is how many runs per prompt you need before a topic number stops moving around on its own. SparkToro listed that as unresolved too. My guess is it depends on category size, since their data showed smaller markets were more stable, though I could be wrong.
If you've turned on Topics Portfolio for your own site, I'd love to hear what it's showing you, especially where it disagrees with whatever paid tool you're using. Send me a note through jarredsmith.com. If you want the longer version of why the engines settle on some brands and keep shuffling others, that's a lot of what my book, Explainable, is about.
Jarred Smith is the author of Explainable: Why AI Recommends Some Brands & Ignores Others, an Amazon bestseller on AEO, GEO, and SEO. He's a marketing leader with nearly 20 years of experience across healthcare, public media, retail, and environmental services. Find him at jarredsmith.com.