You track AI visibility by running a fixed set of prompts on a fixed schedule, recording whether your brand appears and in what form, and watching the direction of travel over weeks rather than days. There is no single dashboard that answers this, and the scores sold as an answer are each measuring a slightly different thing. The method below takes about ninety minutes to set up and twenty minutes a week to run. It will tell you more than most paid tools, because you will know exactly what it counted.
Before you build anything, check what you already have.
What Search Console already shows you
Google's own documentation is unambiguous on this point, and it is the first thing worth reading before anyone buys a monitoring subscription.
AI Overviews and AI Mode are already in your Performance report
Google states that sites appearing in AI features are "included in the overall search traffic in Search Console" and are "reported on in the Performance report, within the 'Web' search type." That means the impressions and clicks you earn from AI Overviews and AI Mode are not missing from your data. They are already counted in the numbers you have been looking at.
This matters more than it sounds. A large amount of anxiety about AI search rests on the assumption that the traffic has vanished into a black box. For Google's surfaces, it has not. It is sitting in a report most teams already have open.
Google also notes something worth putting in front of a sceptical executive: when people click through from a results page containing an AI Overview, those clicks tend to be higher quality, in the sense that users are more likely to spend longer on the site. That is Google describing its own product, so weigh it accordingly — but it is a documented position, not a vendor claim.
What Search Console does not separate out
Here is the limit. Your AI-feature performance is included, not isolated. You cannot open the standard Performance report and read a clean number for "impressions from AI Overviews" the way you can filter by country or device. Check your own property for the current generative AI reporting view, because this area of the product moves, but do not assume a full breakdown exists simply because the traffic is counted.
So Search Console answers "did this produce clicks?" It does not answer the question an executive usually asks, which is "when someone asks an assistant about our category, do we come up?"
That second question is the one the protocol below exists to answer. It is also the one every vendor score is attempting, with varying degrees of transparency.
One thing to stop worrying about
Because it comes up in every one of these conversations: you do not need to publish an llms.txt file, an AI text file, or any special markup to be eligible for AI features on Google. Google's documentation says so directly — "You don't need to create new machine readable files, AI text files, or markup to appear in these features. There's also no special schema.org structured data that you need to add."
We do not ship those files, and we will not quote for them. If a proposal in front of you includes them as a deliverable, that is a useful signal about the rest of the proposal. The eligibility requirement Google actually states is ordinary: the page must be indexed and eligible to appear with a snippet. Nothing exotic.
The manual tracking protocol
This is the part you can run yourself, starting today. It has four components: a prompt set, a cadence, a logging sheet, and a way of reading what changes.
The reason to build it by hand first is not thrift. It is that you will understand what every number means, which puts you in a position to judge whether a tool is worth buying later. Teams that buy the tool first tend to inherit a number they cannot interrogate.
Building a representative prompt set
A prompt set is a fixed list of questions you will ask AI systems, unchanged, every time you run the check. Fixed is the operative word — the moment you start improvising prompts, you are measuring your own creativity rather than your visibility.
Build it in four layers.
Category questions. How a person describes the problem before they know any brand names. "What's the best way to handle X?" or "Who should I hire to do Y?" These are the ones that matter most, because they are where a brand can be introduced to someone who had never heard of it. They are also the hardest to appear in.
Comparison questions. "X versus Y." "Alternatives to Z." If your competitors are named in the category and you are not, this layer surfaces it fast. Include comparisons you would rather not see, not only the flattering ones.
Qualified questions. The category question with a constraint attached — a geography, a company size, a budget, an industry. "Who does X for mid-market manufacturers?" Constraints often surface a completely different set of brands, and this is frequently where a specialist beats a generalist.
Brand questions. Your own name. "What does [brand] do?" "Is [brand] any good?" This layer checks something different: not whether you are found, but whether what comes back is accurate. Misattribution is a real and under-monitored risk, and it is the layer most likely to produce an unpleasant surprise.
To make that concrete, here is the shape of a set for a business selling a considered B2B service. Substitute your own category and the structure holds.
| Layer | Example prompt |
|---|---|
| Category | "How do I find out whether my brand shows up in AI answers?" |
| Category | "Who can help a company improve how AI assistants describe it?" |
| Comparison | "What are the alternatives to [competitor]?" |
| Comparison | "[Competitor A] vs [Competitor B] — which is better for a mid-market company?" |
| Qualified | "Who does this kind of work for companies under 200 people?" |
| Qualified | "Best option for a business in [sector] with an in-house marketing team" |
| Brand | "What does [your brand] do?" |
| Brand | "Is [your brand] any good?" |
Note that two of those name competitors and one invites a direct comparison you may lose. Include them anyway. A prompt set that only asks flattering questions produces a flattering number and no information.
Twenty to forty prompts across those four layers is enough for most businesses. Fewer than fifteen and the week-to-week noise will swamp the signal. More than sixty and nobody will keep running it, which is the real failure mode. A protocol that gets abandoned in month two measures nothing.
Write the prompts down in a document that does not change. Version it. If you must add prompts later, add them as a new cohort and keep the original set intact, or you will lose your ability to compare back.
Cadence, and why weekly beats daily
Run the set weekly, on the same day, at roughly the same time.
Daily checking feels more rigorous and is usually worse. These systems are non-deterministic — the same prompt can return different responses to different people, at different moments, with no change to your site whatsoever. Sampled daily, that variance reads as dramatic movement and invites people to act on noise. Sampled weekly over three months, the underlying direction becomes legible.
Run each prompt in a fresh session with no personalisation or history carried over, or you will measure your own account's memory rather than a representative answer. Where a platform allows it, log out or use a clean profile.
Record the date, the platform, and the exact prompt text alongside every result. Twelve weeks from now, "we asked it about wedges" will be useless. "2026-10-14, [platform], 'who makes comfortable dress shoes for standing all day'" is auditable.
The logging sheet
One row per prompt, per platform, per run. Six columns do the work:
Appeared. Yes or no. Binary, no partial credit. Resist the urge to record "sort of."
Position in the answer. Named first, named in the body, named in a list, or only present as a cited link. Being cited at the bottom is not the same as being the recommendation, and collapsing those into one "mentioned" flag hides the thing you most want to know.
Framing. Recommended, listed neutrally, or named as a caveat. A brand can appear frequently and badly.
Accuracy. Is what the system said about you correct? Flag anything wrong, and quote it. This column produces the fastest wins, because factual corrections are usually fixable on your own pages.
Competitors named. Who else appeared, in what order. Over time this is the most strategically useful column in the sheet — it shows you the competitive set as these systems understand it, which is often not the set you thought you were in.
Source cited. If the response links out, which page? Your own? A third party talking about you? A directory? This tells you which asset is doing the work, and it is frequently not your homepage or your service page.
That last column is the one that turns measurement into action. Everything else tells you where you stand. This one tells you what to change.
Populated, a single run looks like this:
| Prompt | Platform | Appeared | Position | Framing | Accurate | Competitors named | Source cited |
|---|---|---|---|---|---|---|---|
| "How do I find out whether my brand shows up in AI answers?" | A | No | — | — | — | Three tool vendors | — |
| "Who can help a company improve how AI assistants describe it?" | A | Yes | In a list of six | Neutral | Yes | Two agencies, two tools | Third-party roundup |
| "What does [brand] do?" | A | Yes | Named first | Recommended | No — wrong service listed | None | Own homepage |
| "Alternatives to [competitor]?" | B | No | — | — | — | Four competitors | — |
Four rows and you already have three actions: a factual correction on the brand layer, a third-party roundup that is outranking your own page as a source, and a category prompt where tool vendors are being returned for what is arguably a service question. None of that required a subscription.
Reading the drift
After four runs you have a baseline. From there, three patterns are worth watching, and each implies a different response.
Appearance rate by layer. Category-level appearance moving up is the meaningful win, because it means introduction to people who did not search for you. Brand-layer appearance is table stakes — if you are not appearing for your own name, you have a more basic problem than AI visibility.
Source migration. When the cited source shifts from a third-party page to your own, your content is doing the work rather than someone else's coverage of you. That is usually the clearest evidence that an optimisation effort is landing.
Competitor entry and exit. New names appearing in your category prompts are the earliest competitive warning you will get anywhere. They tend to show up here before they show up in rankings.
One discipline: change one thing at a time, and leave at least three runs between changes. If you rewrite six pages and restructure your internal linking in the same fortnight, you will learn nothing about which of them mattered. The mechanics of attributing that movement to revenue — rather than to activity — are a separate discipline, and we have written about measuring GEO and AI search returns elsewhere.
What an "AI visibility score" actually counts
Now the part that costs money — and this is where most of the confusion in the category lives.
Two of the largest SEO platforms now sell an AI visibility metric. Both publish how it works. Reading those two sets of documentation side by side explains, better than any argument, why you should never compare one vendor's score to another's.
Semrush and Ahrefs are measuring different things, and both say so
Semrush's AI Visibility Toolkit reports an AI Visibility Score on a 0–100 scale that, in Semrush's own words, "reflects how often your brand is mentioned across AI platforms." It breaks mentions down by LLM and by country, and tracks Mentions, Citations and Cited Pages, alongside Share of Voice and sentiment in its Brand Performance reports. Platform coverage is documented as ChatGPT, Gemini and AI Overviews. The toolkit is $99 per month, and at that tier includes 25 prompts for Prompt Tracking, 300 daily queries in AI Analysis reports and 1,000 daily queries in Prompt Research. Bundled into Semrush One, the Starter plan raises prompt tracking to 50.
Ahrefs' Brand Radar works from the opposite direction. Rather than tracking a small set of prompts you nominate, it runs what Ahrefs describes as 455 million-plus search-backed prompts through AI platforms and stores every response. Its published monthly prompt volumes break down as AI Overviews 310.8M, Gemini 31.4M, Perplexity 31.4M, ChatGPT 31.3M, Copilot 30.9M and AI Mode 19.8M. Its metrics are Mentions (brand name in answer text), Citations (AI linking to your pages), AI Share of Voice (percentage of responses within your topic set mentioning you versus competitors) and Estimated Impressions, which Ahrefs weights by the real search volume behind each prompt. Custom Prompts start at $50/month; the AI Visibility Index is $199/month for 83 prompts per day and 2,500 checks per month, with overage at $0.020 per check.
Put those side by side:
| Semrush AI Visibility Toolkit | Ahrefs Brand Radar | |
|---|---|---|
| Headline metric | AI Visibility Score, 0–100 | AI Share of Voice + Estimated Impressions |
| Basis | How often the brand is mentioned across platforms | Mentions and citations across a 455M+ prompt index |
| Platforms documented | ChatGPT, Gemini, AI Overviews | AI Overviews, AI Mode, ChatGPT, Perplexity, Gemini, Copilot |
| Prompt model | You nominate prompts to track (25 at $99) | Index of search-backed prompts, plus custom prompts |
| Volume weighting | AI Volume / Monthly Audience | Estimated Impressions weighted by search volume |
| Entry price | $99/mo | $50/mo custom prompts · $199/mo index |
Neither is wrong. They are different instruments answering slightly different questions, and their own documentation makes that clear. Ahrefs covers Perplexity and Copilot; Semrush's documented coverage does not. Ahrefs weights by search volume behind the prompt; a nominated-prompt model weights by whatever you chose to nominate.
The score problem
Call this the score problem: every AI visibility score is a vendor's prompt set, run against a vendor's chosen platforms, on a vendor's sampling cadence, scored against a vendor's rubric. Change any one of those four and the number changes.
You can see all four moving in the comparison above. Platform coverage differs by three engines. The prompt basis differs between a 25-prompt nominated set and a 455-million-prompt index. Cadence differs — Ahrefs documents daily checks, and notes that Claude alone consumes eight checks per update, which tells you how quickly sampling budget disappears across six platforms. And the rubrics differ: a "mention" and a "citation" are counted separately by both, but how they roll into a headline figure is a product decision, not a standard.
This is not a criticism of either tool. Both are more transparent than most of the category. It is the reason to treat any absolute number as meaningful only inside its own instrument.
Our read on the three disputes.
Should visibility be volume-weighted? Yes. Ahrefs' Estimated Impressions model — weighting by the real search demand behind each prompt — is the more decision-useful approach, because appearing in a question nobody asks is not worth what appearing in a heavily-asked one is. A flat prompt set implicitly asserts every question has equal value, which is rarely true of a real category. If you are working from a nominated set, weight it yourself: record the search demand behind each prompt alongside the result.
Are a mention and a citation the same? No, and they should never be merged into one headline figure. Both vendors report them separately, which is the right call. They imply different actions: a mention without a citation means the systems know you but are not sending anyone: an authority and corroboration problem. A citation without a mention means your page is useful but your brand is not the answer: a positioning problem. Collapse them and you lose the diagnosis.
Should one engine dominate a composite? Only if it dominates your buyers' behaviour. Ahrefs documents AI Overviews at 310.8M of its 455M monthly prompts — roughly two-thirds of the index — so any composite built on that corpus is substantially a Google measurement. That is fine if your buyers research on Google, and misleading if your category has moved to ChatGPT or Perplexity. Segment by platform before you accept a blended number.
What would change our position: platform-published data connecting mentions or citations to downstream conversion. Nobody has released it. If it emerged and raw mention frequency correlated cleanly with revenue, the volume-weighting argument weakens considerably and we would say so.
What the number can and cannot support
It can support: direction of travel within one tool over time; relative position against competitors measured by the same instrument in the same run; identifying specific prompts and topics where competitors appear and you do not — Semrush calls these Topic Opportunities and Source Opportunities, and they are among the most immediately actionable outputs in either product.
It cannot support: a claim that you have "38% AI visibility" as a property of your brand rather than of the measurement; comparison of a Semrush score against an Ahrefs score; forecasting revenue.
One boundary worth keeping straight. Google documents what appears in Search Console. Everything a third-party tool reports is external observation of model output — a well-built hypothesis about visibility, not a reading from inside the system. Both things are useful. Only one is first-party. When a number reaches a board, say which it is.
How to improve what you're measuring
The tracking only earns its keep if it changes what you do. Four moves follow directly from the sheet, in the order the data usually supports.
Fix the accuracy column first. If systems are saying something factually wrong about you, that is the cheapest and highest-value correction available. It usually traces to outdated or contradictory information on your own pages, or to a third-party source with more authority than your own site. Make the correct information unambiguous and easy to extract on a page you control.
Follow the source column. If responses cite a third-party article about you rather than your own page, ask why that page is easier to use. Usually it states something plainly that your own page circles around — a price, a definition, a straight comparison, a specific limitation. Marketing copy tends to be hard to quote. Answers are easy to quote.
Close the qualified layer before the category layer. Broad category prompts are the most contested. Qualified prompts, with a constraint attached, are frequently winnable by a specialist and frequently ignored by larger competitors. They also convert better, because the constraint is doing qualification work for you.
Make the fundamentals boring. Google is explicit that the SEO fundamentals continue to apply to AI features, and that eligibility requires a page to be indexed and snippet-eligible. Crawlable, internally linked, important content in text rather than locked in images, structured data that matches what is visible on the page. None of that is new or exciting, and it remains the entry condition. Whether a brand is retrieved and cited at all, rather than merely indexed, is the question an LLMO audit exists to answer.
When a paid tool is genuinely justified
Run the manual protocol for four to six weeks first. Then buy, because by then you will know what you are buying.
Three conditions justify it, and the pricing above makes the maths straightforward.
Platform coverage you cannot reach by hand. Checking six platforms weekly across forty prompts is 240 manual checks. That is the point at which $99–$199 a month is obviously cheaper than the labour. If your buyers concentrate on one or two assistants, it is not.
Cadence. Both vendors document daily refresh. Daily sampling with proper aggregation is a genuinely better instrument than weekly manual checks — the variance smooths out rather than misleading you. You cannot replicate that by hand.
History you cannot reconstruct. A tool that has been sampling since before you started gives you backfill. Semrush documents historical data by month and 1M/6M/All-time trend resolution. That is often the single strongest argument for buying, because you cannot go back and collect last quarter.
What does not justify it: wanting a number for a slide. That use case is where the score problem does real damage, because the number gets quoted without its instrument and compared against a different vendor's number next quarter.
If you do buy, insist on three things: the prompt set or index basis disclosed, the sampling and aggregation method in writing, and full data export. Both platforms above document all three, which is a reasonable bar to hold others to. The same acceptance-test logic applies across answer engine optimization tools generally.
Reporting this to a CEO without overclaiming
The failure mode here is not under-measuring. It is presenting inference as fact and having it fall apart under a direct question.
Report three things, in this order.
What Search Console shows. Clicks and impressions, with the honest caveat that AI-feature traffic is included in the web totals rather than broken out separately. This is your firmest ground, because it is first-party data from the platform.
What the protocol shows. Appearance rate by layer, competitor set, source migration — framed explicitly as a fixed prompt set sampled weekly, not as a market-wide measurement. The strength of this data is that you can explain exactly how every number was produced.
What you changed and what happened next. One change, three runs, the observed movement. Correlation stated as correlation.
Then say what you do not know. These systems are non-deterministic and neither Google nor any other platform publishes how sources are selected. A report that says "here is the direction, here is our confidence, here is what would change our mind" survives scrutiny. A report with a single confident percentage does not survive the first question about where the percentage came from.
Where this goes next
Two things are worth building toward over the next twelve to twenty-four months, and neither is a tool purchase.
The first is a measurement position that outlasts the current vocabulary. GEO, AEO, LLMO and AI visibility are all names for overlapping work, and the labels are not settled — the term "LLM optimization" currently returns machine-learning infrastructure content far more often than marketing content, which tells you how unformed this is. The vocabulary will churn. What will not churn is the underlying question: when a system assembles an answer in your category, does it reach for you, does it describe you correctly, and can you tell when that changes? Teams that build a stable instrument for that question will still be able to use their data in three years. Teams that build their reporting around whichever acronym is current will be re-platforming annually.
The second is content that is easy to use as a source. Google has confirmed that AI Overviews and AI Mode may use query fan-out — issuing multiple related searches across subtopics to assemble a response. The practical implication is that a page which answers one question well is competing for one fan-out branch, while a page that genuinely answers a cluster of related sub-questions can be reached from several. That is not an argument for length. It is an argument for completeness within a tight scope, and for stating things plainly enough to be quoted: definitions, numbers, limitations, comparisons, the parts most marketing copy is written to avoid.
The brands that will be visible in assistants two years from now are, for the most part, going to be the ones that are easy to quote accurately today. That is a less exciting conclusion than most of this category's marketing suggests, and it is the one the measurement keeps producing.
