Start with the scope, not the score
An AI visibility report records how a defined set of AI assistants answered a defined set of questions during a defined period. Read those inputs first. Then inspect the mentions, citations, competitors, and any ordered positions. Treat the headline score as a summary of that sample, not as your position across AI search.
That distinction matters because AI answers vary. Two identical questions can return different brands and sources, and changing the prompt, assistant, market, or date changes what you measured.
The IAB's August 2026 AI visibility measurement framework separates directional evidence from measurement intended for larger business decisions. Its practical message for a small business is simple: decide what the report is good enough to tell you before acting on it.
1. Check what the report actually covered
Look for the exact assistants, model or mode where available, target market, collection dates, prompt count, and number of completed answers. A report about ChatGPT, Gemini, and Claude does not establish what happened in Perplexity or Google AI Overviews.
Keep each assistant visible rather than relying only on a combined figure. In a Semrush study published in 2026, the tested engines disagreed about whether to name a brand in 100 of 454 prompt-and-domain comparisons. That is a result from one defined dataset, not a universal disagreement rate, but it shows why assistant-level evidence matters.
2. Read the questions before the results
The question set defines the report. A set full of branded questions such as “What does Meridian Accounting offer?” will produce more mentions than a set of discovery questions such as “Which accountants help Austin software startups with R&D tax credits?” Those sets answer different questions.
Check that the prompts reflect real customer situations, cover more than one buying need, and separate branded questions from questions where the customer has not chosen a provider. Keep the wording stable when comparing two reports. Our guide to choosing AI visibility prompts gives you a practical review process.
The August 2026 IAB framework goes further for category-level measurement. It treats programs with fewer than 50 distinct queries as exploratory and asks providers to disclose how prompts are selected, grouped, weighted, and refreshed. Mentsight uses a smaller set of 10–15 owner-approved questions and repeats each one five times per assistant. That makes a Mentsight report a focused website snapshot, not a complete measurement of a market.
3. Check the answer count and the variation
A result based on one response should be labelled as one response. Repeating the same question gives you a better view of normal variation, but it does not remove uncertainty.
Two 2026 research preprints make this limitation concrete. One study of citation variability repeatedly sampled Perplexity Search, OpenAI SearchGPT, and Google Gemini and found substantial variation in cited domains. It concluded that single-run figures can look more precise than the evidence supports. A second preprint likewise recommends treating visibility as a distribution rather than one fixed outcome.
In a Mentsight report, questions are asked in five rounds per assistant. Failed answers reduce the usable basis rather than becoming false “not mentioned” results. Read the completed-answer denominator and the reported 95% range beside the score. A movement from 61 to 64 means little if the two supported ranges overlap.
4. Keep mentions, citations, and positions separate
Mentioned: the business name appeared in the answer.
Cited: a returned source link pointed to the business's website.
Position: the business had a place in an explicitly ordered recommendation.
These observations can happen independently. A page may be cited as supporting material while the answer never names the business. A business may be mentioned without its website being cited. A narrative answer may mention several businesses without ranking any of them.
The Semrush study logged 3,981 domain appearances across 115 prompts and four AI search experiences. In that sample, 61.7% of appearances were citations without a corresponding brand mention. The useful conclusion is not that every category will show the same percentage. It is that citation rate and mention rate describe different things and should remain visible separately.
5. Unpack the headline score
Before using a visibility score, find its formula and denominator. Check which metrics are included, how they are weighted, whether branded questions count, how failed answers are handled, and whether assistants are pooled.
Our AI visibility is 64 out of 100.
This report's score is 64/100, based on its eligible completed answers and the stated mention and owned-citation weights.
Mentsight's visibility score summarizes one report on a 0–100 scale. Being named supplies 70% of the score and having an owned page cited supplies 30%. Questions that name the tracked business are excluded. Position is not scored because many useful answers are not ordered lists. The score cannot be compared with a score from another tool that uses a different question set or formula.
The IAB framework makes the same transparency point more broadly: a composite index is a presentation layer. Its component values, weighting, normalization, and limits should remain available. The Mentsight measurement method documents those choices for our reports.
6. Open the evidence behind an interesting result
The original answer is where interpretation starts. Open it and check:
- The question and assistant
Confirm the exact wording and which assistant produced the response.
- The name in the answer
Check that the answer actually names the business.
- The destination of each source
Separate links to the business's own domain from directories, publications, and other external sources.
- The claim each source supports
Open important sources and read the relevant page in context.
- The shape of the recommendation
Only treat a number as Position when the answer genuinely orders options.
OpenAI's current ChatGPT Search guidance says search results and citations may be incomplete, outdated, or incorrect and advises readers to open important sources. A citation flag alone therefore is not enough evidence for a website change.
7. Compare like with like
For a useful comparison, keep the question wording, assistant set, market, and counting method stable. Record model or platform changes when they are known. Then compare the supported ranges and the underlying answers, not only the two headline numbers.
Google's documentation for AI features in Search says AI Overviews and AI Mode can use different models and techniques, so their responses and links can vary. It also reports traffic from those features inside the general Web search type in Search Console. That traffic can help you understand downstream visits, but it does not provide the same prompt-and-answer record as a visibility report.
8. Turn one pattern into a bounded next step
Use the report to choose what to inspect, not to declare what caused the result. In the illustrative Mentsight report, Meridian Accounting appears in 12 of 16 answers and its own site is cited in six. It appears most often for service questions and not at all for the sample's price-led questions.
That pattern does not prove Meridian needs a pricing page or that publishing one would change a future answer. It gives the fictional owner a useful check: do customers need clearer pricing or fee guidance, and does the existing site answer that need accurately?
Illustrative example: Meridian Accounting is fictional, and the numbers come from Mentsight's public sample report. They are not a customer result or an industry benchmark.
The AMEC practitioner guide to generative engine optimisation measurement says there is no single stable metric for total visibility and that a changed AI answer does not prove business impact. Where links exist, use analytics to check referral sessions and conversions. Combine that evidence with the report before making a larger decision.
A checklist for the next report you read
I can see the exact prompts, assistants, market, dates, and completed-answer count.
I know whether the prompts are branded or non-branded and why they were chosen.
Repeated answers and uncertainty are visible.
Mentions, owned citations, external sources, and ordered positions stay separate.
The score's formula, weights, exclusions, and denominator are explained.
I can open the full answer and returned sources.
Comparisons use the same measurement basis or clearly disclose what changed.
The next step is framed as something to check, not a guaranteed improvement.
If any of those details is missing, narrow the conclusion. The report may still be useful as a starting point, but it cannot support a claim its evidence does not reach.
Source review: research reviewed 2 September 2026. Editorial owner: Mentsight. Review again by 2 December 2026 because platform documentation and measurement standards can change.