Understanding Your AI Visibility Report
An AI visibility report shows how AI tools describe your business when people ask questions about what you sell or do. It looks at whether your brand appears in those answers, whether what they say about you is accurate, and whether the answer links people to your site.
We use the first report as a baseline and repeat it quarterly or every six months, depending on the work underway and how quickly your market is changing. The report is most useful for established brands with enough search presence, content and competition to measure. For a newer or smaller business it can still be a useful starting point, though it will usually show more opportunity than activity.
You receive two versions. The brief is the short report with the main findings, priorities and next steps. The technical report is the supporting work behind it, with every question, answer, scoring rule, source and limitation. Most clients only read the brief. This page explains the brief first; the method notes at the end cover the technical report.
One rule applies throughout. We keep recorded business results separate from sampled AI-answer data. A visit or a sale recorded in your own analytics is not the same kind of evidence as an AI tool mentioning your brand in a test answer, so we never blend the two into one score.
What the Report Can and Cannot Tell You
The report can tell you how often a fixed set of AI questions named your business, recommended it, or linked to your site, and how that compares with a fixed group of competitors on the day of the test. It can tell you what the AI tool says about you and which sources it built that picture from. And it can show which of your topics already carry search demand and revenue, and who the AI tools cite most on each.
It cannot tell you how many people asked those questions, how many saw a given answer, or how many bought something because of one. AI answers change from one run to the next, so a single result proves little. And no tool covers every assistant. The report says which ones were tested.
Two Kinds of Evidence
Every number in the report carries one of two labels.
| Label | What it means | Examples |
|---|---|---|
| Recorded | Activity measured by your own analytics and Google tools | Search Console clicks, Google Analytics sessions and recorded revenue |
| Sampled | A structured review of how AI tools present your business beside competitors | Mentions, links to your site, citation share, and prompt-based comparisons |
Recorded data tells us what happened on your website. Sampled data tells us how often a brand appeared in a defined set of AI answers at a point in time.
Neither is complete on its own. A sampled AI answer cannot prove traffic or revenue. Recorded traffic cannot show the times an assistant mentioned your brand without sending a click. Keeping them apart is what makes the report both honest and useful.
How We Test AI Answers
AI answers are not fixed. Ask the same question twice and you may get different wording, different sources or different recommendations.
To account for that, we use a fixed set of questions, usually 50 to 100, and ask each one three times on two versions of the AI model. The same set is used again in later reports, so we compare like with like over time instead of relying on one answer from one day. The brief states the question count, the number of runs, and the total answers behind it.
The questions cover different moments in a buyer’s decision.
| Question type | Example | What it tells us |
|---|---|---|
| Brand questions | ”Is [brand] reputable?” | What the AI tool says about your business and where that information came from |
| Comparison questions | ”Is [brand] or [competitor] better for…?” | Whether you appear when shoppers compare options |
| Discovery questions | ”What are the best [products] for…?” | Whether you are found when no company is named |
| Purchase questions | ”Where can I buy…?” | Whether you are recommended when intent is strongest, and where competitors most often win the click |
| Help and safety questions | ”How should I use…?” or “Is this safe for…?” | Whether you appear in education-led or safety-sensitive answers. These often go to professional bodies and reference sites, so low numbers here are normal |
Brand-named questions are kept out of the discovery and purchase rates. A question that already includes your company name would make you look more visible than you are.
The competitors are a fixed list with a version date. Changing the list resets the trend, so the report says which version each run used.
What Mentioned, Recommended and Linked Mean
The brief focuses on three results for each question type, for your business and for the average competitor.
| Result | What it means | Why it matters |
|---|---|---|
| Mentioned | The answer names your business | The AI tool sees you as relevant to the question |
| Recommended | The answer presents your business as an option to buy from or use | Stronger than a passing mention. It suggests fit or preference |
| Linked to your site | The answer links to a page you own | The result most likely to send someone to your website |
A business can be mentioned often and linked rarely. That does not by itself mean something is wrong, but it is a useful clue. It often means the AI tool relies on reviews, marketplaces, forums or publishers when it talks about you rather than on your own pages. The brand-answer section shows what the tool credits you for, what it warns shoppers about, and which sources it drew on, which is why the report also checks whether the tool can read your shipping, returns and quality pages at all.
The competitor comparison is not an industry benchmark. It shows how you appear within the specific set of alternatives a buyer is likely to meet.
What the Scorecard Shows
The technical report opens with five measures, each with your rank among the compared businesses and the change since the last run. They are reported separately and never combined.
| Measure | What it shows | Label |
|---|---|---|
| AI citation share | Of the Google searches where an AI Overview cites you or a compared competitor, the share that cite you | Sampled |
| Prompt mention rate | The share of our unbranded test questions where ChatGPT named your business, with recommendation and link rates beside it | Sampled |
| Organic demand | Google clicks from searches that do not include your brand name, over the last 90 days against the 90 before | Recorded |
| Authority context | Domain Rating and organic keyword counts, as context for a citation gap | Sampled |
| Business impact | Visits your analytics attributes to AI assistants, their share of all visits, and the revenue from them | Recorded |
Organic demand is on the scorecard because Google’s AI features draw on the same pages Google ranks, and strong search performance often supports eligibility to appear in them. It does not guarantee any particular AI result, which is why the two are reported apart.
How to Read the Topic and Page Findings
The brief picks a handful of product or service topics that already carry search demand and shows, for each, the recorded clicks and revenue on those pages and which compared business Google’s AI Overviews cite most.
| Column | What it means | Label |
|---|---|---|
| Non-brand Google clicks, last 90 days | Clicks from Google Search to your pages on that topic, brand-name searches removed, against the 90 days before | Recorded |
| Revenue from visits to these pages | Revenue from every visit that started on those pages, from any source. It shows what the pages contribute today, not revenue caused by AI | Recorded, reported as a minimum where cookie consent limits what Analytics records |
| Share of AI Overview citations | Of the times Google’s AI Overview cites one of the compared businesses on the topic, the share that cite you | Sampled |
| Most-cited | The compared business with the largest citation share on the topic | Sampled |
Each topic carries one of two labels. Protect and extend means you already lead the topic and the work is keeping that position. Main growth work means a competitor leads, and the report names them so you can see whose pages the AI tools prefer.
The technical report adds a table of every page of yours an AI source cited, joined to that page’s clicks and revenue, and a second table of pages that bring in orders or inquiries but were not cited in any sample. Not cited in the sample does not mean never cited. It means the page needs a look, because it earns money and no AI tool in the test used it.
Why the Sources Sometimes Disagree
The report draws on your own Google Analytics and Search Console, on our ChatGPT question panel, and on two third-party datasets of AI citations. Different tools sample different questions and sources, so they will not always agree. A brand can rank second in one dataset and fifth in another on the same day, and both can be right about their own sample.
When one source places you above the median and another places you below for the same AI tool, the report shows both and flags the disagreement instead of picking one. A favorable result from one source does not cancel out a weaker result from another.
What Google’s Own Tools Measure
Google Search Console now includes a Generative AI performance report for AI Overviews and AI Mode. It shows how often links to your site appeared in those features, by page, country and device, and Google rolled it out to every property on August 31, 2026. It shows impressions, not clicks, and Google still counts the clicks inside ordinary web search.
Google Analytics can identify some visits from AI assistants through its AI Assistant channel, added in May 2026. Some AI-driven visits arrive without a recognizable referrer, and visits from Google’s own AI features are still recorded as ordinary Google organic traffic. We treat reported AI-assistant visits as a documented minimum, not a complete count.
Neither tool shows how ChatGPT, Perplexity, Gemini or Copilot describe your business. That is what the question panel and the third-party datasets are for.
What Happens After the Report
The brief ends with a short list of actions for the next 90 days. Each has an owner, a reason, and a way to confirm whether it worked.
Crawler Access Is Checked First
If an AI search crawler cannot read a page, that page may not be eligible to appear in that tool’s results or citations, so we check access before recommending content work.
A simple test can mislead, especially on a site behind a CDN, firewall or bot-management service. A copied crawler user agent that receives a 403 does not prove the real crawler is blocked. When access matters, we confirm it through server or CDN logs and the platform’s own tools. Allowing a crawler does not guarantee that an AI tool will cite or recommend a page.
Typical Actions in the 90-Day List
- Confirming that AI crawlers can reach the site, before any content work
- Correcting inaccurate information about the business on high-visibility third-party sites
- Improving product, service, shipping, returns and policy pages so the facts are easy to find
- Strengthening the pages competitors are cited for more often
- Reviewing pages that bring in orders or inquiries but were not cited in the sample
We do not judge success by one mention or one link in a later answer. On the next run we ask the same questions, on the same models, against the same competitors, and look for mentions, recommendations and links moving in the same direction over at least 30 days, alongside recorded changes in search clicks and revenue on the pages we changed. A first report is a baseline and says so.
Want the Supporting Detail?
Most clients only need the brief. If you want to review the underlying questions, answers, scoring rules, source limitations or raw figures, ask for the technical report. AI visibility is still a new kind of measurement, and the right response to a number you do not understand is to check what it represents before acting on it.
Questions about a specific figure can go to the same email address that sent your monthly marketing report.
Method Notes for the Technical Report
Everything below is for readers of the technical report, or for anyone who wants to check the work. None of it is needed to read the brief.
How each answer is scored
Every answer is scored by fixed rules with a version number. The rules do not change between runs without the version changing, which is what lets us read a change in the numbers as a change in the answers.
| Term | Rule | Why it is defined this way |
|---|---|---|
| Named | The brand name appears anywhere in the answer | The loosest test. Awareness |
| Recommended | The answer suggests the brand as a place to buy or use. Ambiguous wording is marked unknown, and unknown is never counted as no | Counting an ambiguous answer as a miss would understate every brand equally and hide real differences |
| Owned site cited | The answer links to a page on the brand’s own domain, matched on the host | A review site’s page about the brand does not count. Only your own page can send a visitor to you |
| Share of voice | The brand’s mentions divided by all mentions of the brand and its compared competitors in the same answers | Publishers, forums and marketplaces are excluded, so the share is among businesses that sell what you sell |
| Prominence | 0 absent, 1 passing mention, 2 named in a list, 3 secondary recommendation, 4 primary recommendation, 5 primary recommendation with a link to the brand’s own site | Being tenth in a list and being the first recommendation are not the same result |
| Score | Prominence points earned divided by points possible, as a percentage | One figure per group that rewards being recommended over being mentioned |
A row built on fewer than 30 answers is marked n<30 and treated as sampled context only. The technical report also carries a directional index, a single 0 to 100 figure built from the question-type scores with fixed weights. It is for our internal sorting across clients and never appears on the scorecard, because it is a judgment about which question types matter, not a measurement.
How questions are tagged
Each question carries five tags so answers can be grouped several ways. Question type is the grouping the brief uses (brand, comparison, discovery, purchase, help and safety). Stage records how close to buying the person is (discovery, consideration, fit help, purchase). Persona records the kind of customer (beginner, budget, premium, and others specific to the client). Topic cluster records which product or service category the question belongs to. Source records who wrote the question, so we can check whether the questions the client suggested behave differently from the ones we wrote. Brand-named questions are held to about a quarter of the panel or less.
What each source can and cannot see
| Source | Can see | Cannot see |
|---|---|---|
| Google Analytics 4 | Visits whose referrer is an AI assistant, or that GA4 placed in its AI Assistant channel (from May 13, 2026; earlier months are matched by referring domain), with engagement, purchases and revenue | Answers that never send a click. Visits without a referrer, which land in Direct. Clicks from Google AI Overviews and AI Mode, reported as ordinary Google organic traffic |
| Google Search Console | Clicks and impressions for Google web search by query and page, and, in the Generative AI performance report, impressions in AI Overviews and AI Mode by page, country, device and date | Clicks or queries from AI Overviews and AI Mode. The report we build from the Search Console API does not yet include the generative report’s figures, so they are read by hand |
| Ahrefs | AI citations per domain by platform, a custom Brand Radar report on a fixed prompt list, Domain Rating and organic keyword counts | Answers outside Ahrefs’ sampled prompts. Brand Radar share of voice is weighted by how often each prompt is asked, so branded prompts carry most of it and shares do not sum to 100 percent |
| DataForSEO | Google AI Overview citations per keyword, how many answers in its ChatGPT and AI Overview index cite or name each brand, and live ChatGPT answers to our panel | Assistants other than ChatGPT and Google. How often real people ask the panel questions |
| Crawler access check | robots.txt rules per crawler and how the site answered each user agent on the day of the check | Whether the real crawlers can fetch from their own IP ranges |
The tools use different question sets, so the report compares down a column, never across.
Search Console proxies
Beside the Search Console figures, the technical report tracks the share of clicks from queries of eight words or more and from question-style queries. Longer and question-style searches are context for how people phrase information-seeking searches. They do not identify AI-driven behavior on their own, so the report treats them as supporting context rather than proof. Impressions are collected but kept off the page, because Google has acknowledged inflated impression counts.
How clusters and cited pages are built
AI Overview citations come from each compared domain’s top 1,000 cited keywords by search volume, assigned to a topic cluster by query pattern. Search Console clicks are matched to a cluster the same way. Analytics sessions and revenue count every visit that started on a page in the cluster, from any channel, with a separate column for visits from AI assistants. Cited page URLs are normalized before the join (tracking parameters removed, page-defining parameters kept), because the same page cited three ways is one page.
Breaks in the data
Analytics data changes meaning when the measurement setup changes, and the report marks those dates on every chart. The common one is a consent banner switched to opt-in. From that day, Google Analytics records only visitors who accept cookies, so sessions and revenue drop by roughly half to two thirds with no change in the business. The report tags each month as before, mixed or after, compares shares within Analytics rather than counts across the change, and states the assumption that carries the comparison: that visitors arriving from AI assistants accept cookies at about the same rate as everyone else. That assumption is unverified, and the report says so. Other marked events include the date the GA4 AI Assistant channel started and any move between hosting or CDN providers.
Which crawlers matter
OAI-SearchBot decides whether a page can appear and be cited in ChatGPT search. ChatGPT-User fetches a page live when an answer needs it. GPTBot is OpenAI’s training crawler, and blocking it does not affect search citations. Bing’s crawler matters for ChatGPT too, because ChatGPT search draws on third-party search providers and Bing is the one that has been named. Googlebot covers Google Search, AI Overviews and AI Mode together. Confirmation that a real crawler gets through comes from CDN or firewall logs filtered to each crawler’s published IP ranges, from Search Console’s URL Inspection for Googlebot, and from Bing Webmaster Tools for bingbot.
Corrections
September 24, 2026. The first version of this page, and the first September technical report, said Search Console could not separate AI Overviews and AI Mode and had no generative report. Google rolled the Generative AI performance report out to all properties on August 31, 2026. The claim had been checked against a developer page that had not been updated rather than the Search Console help center. Both were corrected the same day, and rebuilt reports carry the correction in their opening notes.