Leadership teams increasingly ask a reasonable question: when people ask an AI assistant about our category, do we appear, and is what it says accurate? The honest answer is that this can be assessed, but not measured with the precision people are used to from traditional search reporting. Answers vary between runs, users, locations and product versions. Several products publish little or no data about how often a brand is shown.
The thesis of this article is that AI search visibility should be assessed through a documented, repeatable sampling method, reported as observations with stated limits, and never as a single score presented as fact. A careful baseline is useful. An overconfident one is worse than none, because it leads to decisions based on noise.
This work is often called GEO (Generative Engine Optimization) or AEO (Answer Engine Optimization): helping AI assistants and answer engines find, understand and cite your content. Neither label changes the measurement problem, so a GEO programme needs the same honest baseline before anyone claims it is working.
What follows is a practical method: what to observe on each surface, how to design a prompt set, how to record and score results transparently, and which caveats belong in every report.
Start with what can actually be observed
Different AI-assisted discovery products expose different information. Treating them as one channel hides important differences. The table below describes the general situation at the time of writing; products change quickly, so verify current behavior before relying on any row.
| Surface | What you can usually observe | What is usually hard or impossible to observe |
|---|---|---|
| Google AI Overviews (in Search results) | Whether an overview appears for a query in a given session, its text, and the links shown alongside it | How often it appears across all users; a separate, complete performance breakdown for AI features in standard reporting |
| Google AI Mode | The conversational answer and linked sources for a given session | Population-level frequency of mentions; how follow-up context changes answers for other users |
| ChatGPT and similar general assistants | The answer for a given prompt, account state and model; source links when the product searches the web | Whether other users receive similar answers; how memory, settings or model versions affect results; impressions of any kind |
| Other assistants and answer engines | Varies by product: some show sources consistently, some rarely | Usage volumes and any form of impression data |
| Your own analytics | Some referral visits from AI products, where a referrer is passed | Visits where the referrer is stripped; influence on users who never click |
Where a brand can be observed in AI-assisted discovery
Your brand at the centre, observed through four kinds of surface. Each shows something different: AI features in Google Search, general assistants, other answer engines, and your own analytics, which only sees attributable clicks.
Your brand
as it appears in sampled answers
Google AI features
Per session
- AI Overviews and AI Mode answers
- Links shown alongside them
General assistants
Per prompt
- Answers by prompt, account and model
- Sources when the product searches
Other answer engines
Varies
- Source display differs by product
- No impression data
Your own analytics
Clicks only
- Referral visits where passed
- Misses unclicked influence
Observations are samples of answers, not a census of what all users see.
Google states that visits from AI features such as AI Overviews and AI Mode are counted within the overall “Web” search data of the Search Console Performance report, not as a separate report (Google Search Central: AI features and your website).
Two practical consequences follow. First, any visibility assessment is a sample of observed answers, not a census. Second, traffic data from analytics shows only the part of AI-assisted discovery that results in an attributable click, which is likely to understate influence and cannot be used to infer how often a brand is mentioned.
Design the prompt set
The prompt set is the foundation of the assessment. It should reflect how real buyers ask questions, not how the marketing team describes the company.
Prompt categories
Group prompts so that results can be interpreted by intent:
- Brand prompts — direct questions about the company (“What does [brand] do?”, “Is [brand] reliable for…?”). These test accuracy more than visibility.
- Category prompts — questions where the brand could reasonably be mentioned (“Which providers offer…”, “What should I look for in…”).
- Comparison prompts — the brand alongside named alternatives.
- Problem prompts — the buyer’s problem described without naming a solution category.
Sources for realistic wording include sales call notes, support tickets, on-site search logs and the queries that already bring organic search traffic. Write prompts in the language and phrasing of the target market, and keep the final wording fixed so it can be rerun.
Record the conditions
For every run, record the conditions that can affect the answer:
- Product and mode (for example AI Overview in Search, AI Mode, a named assistant with web search on or off)
- Model or version where the product discloses it
- Date and time
- Account state: logged out, logged in to a clean test account, or logged in to a personal account (avoid the last)
- Location and language settings
- Whether the prompt was the first in a fresh conversation
Without these, results cannot be compared over time, and an apparent change may simply reflect different conditions.
Repeat runs
Because answers vary, run each prompt more than once under the same conditions, ideally at different times. The number of repeats is a trade-off between cost and confidence; whatever number is chosen, state it in the report. A brand that appears in one of several runs is in a different position from one that appears in all of them, and a single run cannot tell the two apart.
Score observations transparently
Store the full text of every answer, with its conditions, before scoring anything. Then score against simple, written definitions so that another reviewer could reach the same result.
| Measure | Definition |
|---|---|
| Mention | The brand name, or an unambiguous reference to it, appears in the answer text |
| Linked source | A page on the brand’s own domain is shown as a source or link for the answer |
| Third-party source | A page on another domain that discusses the brand is shown as a source |
| Accuracy | Each factual statement about the brand is marked correct, outdated, incorrect or unverifiable |
| Position | Whether the brand is named in the opening part of the answer, later, or only in sources |
| Framing | Neutral, positive or negative description, marked by a reviewer with the reasoning noted |
From these, a simple observed mention rate can be calculated per prompt and per category: runs in which the brand was mentioned divided by total runs. Report it with the number of runs alongside it, and avoid combining surfaces into one blended figure, because they are not comparable.
Accuracy findings are often the most actionable output. An assistant that consistently describes a discontinued product, an old price model or a wrong location points to specific information that needs correcting or clarifying across sources the brand controls or can influence.
From observations to a visibility gap
Once a baseline exists, compare it with what a fair answer would contain. For category and problem prompts, ask: which sources are being cited, does the brand have content that genuinely answers the question, and is that content accessible and clear?
Common gaps include:
- Eligibility and access. Important pages blocked from crawling, rendered only with client-side scripts, or behind forms.
- Clarity. Key facts (what the company does, for whom, where, at what scope) are scattered or inconsistent across the website and profiles.
- Third-party presence. Answers draw on reviews, directories, comparisons or publications where the brand is absent or described inconsistently.
- Content fit. No page directly answers the questions buyers ask; the content describes the company instead of addressing the problem.
Structured data can help search engines interpret page content, and it is reasonable to implement accurately where it fits. It does not guarantee inclusion in any AI-generated answer, and no reputable assessment should present it as doing so.
Our AI search visibility work follows this sequence: baseline method, observed gap, content and distribution recommendations, then a monitoring plan using the same prompt set.
Monitoring over time
A baseline is useful mainly as a reference for later comparison. To keep comparisons valid:
- Rerun the same prompt set under the same documented conditions at a regular interval.
- Note product changes you are aware of, such as new modes or model versions, alongside results.
- Add new prompts as a separate set rather than mixing them into the baseline.
- Review accuracy findings first; they are the clearest signal of whether corrections are being picked up.
- Look at trends across many prompts rather than reacting to movement in any single answer.
Caveats
AI-generated answers are non-deterministic and personalized. The same prompt can produce different answers for different people, or for the same person minutes apart. All findings are observations from a sample, not measurements of what every user sees.
Products, models and interfaces change frequently and often without public notice. A change in results may reflect a product update rather than anything the brand did.
There is no reliable public data on impressions or mention frequency for most AI assistants. Any tool or report claiming a precise “AI visibility share” should explain exactly how it was sampled.
Referral traffic from AI products is only partly attributable and is likely to understate influence. It should not be used as a proxy for visibility.
No method or markup can guarantee that a brand will be mentioned or cited in AI answers.
Action checklist
- Decide which AI surfaces matter for your buyers and assess them separately
- Build a prompt set from real buyer language, grouped by intent
- Fix the wording and document run conditions for every observation
- Run each prompt several times and record the number of runs
- Store full answer text before scoring
- Score mentions, sources, accuracy, position and framing against written definitions
- List inaccurate or outdated statements and trace where they may originate
- Check that key pages are crawlable, accessible and clearly worded
- Review third-party sources that answers rely on
- Set a monitoring interval and keep the baseline prompt set unchanged
If you want an independent baseline before investing in AI visibility work, you can request an assessment. We will be explicit about what we observed, how often, and what we could not measure.