How Do You Measure AI Search Visibility?
By Jordan Baker, Founder of Fair Public Relations and former UK national journalist
Appearing in ChatGPT once is exciting.
I know because we have done exactly what most business owners do.
You type a commercially interesting question into ChatGPT.
Your company appears.
You take a screenshot.
You send it to somebody.
And for a moment it feels as though you have cracked the entire future of search.
The problem is that one screenshot tells you very little.
Would the same business appear tomorrow?
Would it appear if the wording changed slightly?
Would it appear for somebody in another country?
Was the company actually recommended, or was one of its web pages merely cited?
Was the description correct?
Did anybody click?
Did anybody enquire?
Did anybody buy anything?
Those are completely different questions.
This is why I do not think AI visibility should be reduced to one invented "GEO score".
A useful measurement system needs to distinguish citations, brand mentions, recommendations, accuracy, referral traffic and genuine commercial outcomes.
Some of those can now be measured directly.
Others can only be observed.
And some remain extremely difficult to attribute confidently.
What does AI-search visibility actually mean?
I find it useful to start with seven different outcomes.
A source citation happens when an AI-generated answer uses or links to one of your pages as supporting material.
A brand mention happens when the answer actually names your company.
A recommendation is stronger commercially. The system suggests your business as a suitable option for the person asking the question.
Accuracy is different again. Does the AI describe the company correctly? Does it understand what you sell, where you operate, your pricing, credentials and limitations?
Then there is visibility without a click. Somebody may see your business in an answer and never visit your website.
There is referral traffic, where somebody does click through from the AI system.
Finally there is the outcome most businesses ultimately care about: commercial action.
An enquiry.
A booking.
A call.
A purchase.
A proposal.
A customer.
These outcomes belong in the same measurement picture.
They should not be treated as the same metric.
Why a citation and a recommendation are not the same thing
Imagine somebody asks:
"What should I look for when choosing a PR agency?"
ChatGPT uses an article from your website as one source but never mentions your agency.
That is a citation.
Now imagine somebody asks:
"Which small-business PR agencies in the UK have transparent pricing?"
and the system names your agency.
That is a brand mention and potentially a recommendation.
Commercially, I care much more about the second.
The first still tells me the site's material is retrievable and useful enough to support an answer.
That is interesting.
But I would never report the citation as though a buyer had just been told to hire us.
Why AI visibility is harder to measure than a conventional Google ranking
Traditional SEO reporting can still be messy, but the basic concept is familiar.
You search a query.
A page appears somewhere in the results.
Search Console gives you performance data.
AI answers are much more dynamic.
Prompt wording can change the result.
Geography can matter.
Context may matter.
Different systems use different sources.
The same platform can answer a question differently over time.
An answer may cite your page without naming the company.
It can name the company without linking to it.
It can recommend a competitor while using your article as supporting material.
And the user can receive enough information from the generated answer that they never click anywhere.
This is precisely why Microsoft describes AI citation data as observational rather than a traditional ranking system. Its Citation Share metric, for example, shows a site's share of citations for particular grounding queries, but Microsoft explicitly says it should not be treated as a ranking or competitive quality score.
That is the right mindset.
Observe patterns.
Do not manufacture certainty.
Start with the questions your real customers ask
The most useful AI visibility test begins before you open ChatGPT.
Ask:
What would a genuinely suitable customer ask?
Not:
"What is the greatest PR agency on Earth?"
That is a vanity prompt.
For Fair PR, we are much more interested in questions around things such as:
PR agencies suitable for small businesses;
PR agencies working with particular sectors;
journalist-led PR;
PR for experts;
boutique agencies;
transparent PR pricing;
and comparisons between agency options.
Those reflect real commercial intent.
They also align with actual parts of our service.
A technology company should develop its own set.
A local restaurant's prompt universe will be completely different.
So will a surgeon's.
The purpose of the test is not to find wording that makes your company appear.
It is to approximate genuine customer discovery.
Commercial fit matters as much as visibility
Suppose Fair appeared in 500 answers to:
"How do I write a university PR essay?"
That might increase an AI-visibility metric.
It would have almost no commercial relevance to us.
Now suppose Fair is repeatedly recommended for:
"PR agency for a UK founder with a £1,000 monthly budget."
That is vastly more interesting.
Fewer appearances can be worth more if they occur at the right point in the buying journey.
This is why I think businesses need to resist the temptation to turn GEO into another impressions race.
How should you establish an AI visibility baseline?
I would start with a defined set of prompts.
Keep a record of the exact wording.
Record the intended market.
Record the platform being tested.
Record the date.
Record whether the business was recommended, mentioned, merely cited or absent.
If it appeared, record which page seemed to be supporting it where identifiable.
Check whether the description was accurate.
Then repeat the exercise.
The key is consistency.
If your baseline test uses:
"Best PR agency UK"
and next month's test uses:
"Which affordable journalist-led PR agency should I use in London?"
you are not comparing exactly the same thing.
Both questions may be useful.
They simply belong as separate measurements.
Do not treat manual prompt testing as laboratory science
Even consistent prompts have limitations.
AI systems change.
Answers can vary.
Personalisation and location may affect some experiences.
A model may search the web for one answer but not another.
Sources can change between runs.
This is why I use prompt monitoring as directional evidence.
If Fair repeatedly appears for a group of relevant commercial prompts over several months, that becomes a much more useful observation than one isolated screenshot.
Can you track AI citations directly?
Increasingly, yes.
This is one of the biggest changes during 2026.
Microsoft's Bing Webmaster Tools launched an AI Performance report in February.
It can show total citations from supported AI experiences, average cited pages, page-level citation activity, visibility trends and samples of the grounding queries used when content was retrieved. Microsoft makes an important distinction: a citation count tells you that content was used as a source, not that the page held a particular "rank" within an answer.
Then in June, Microsoft added Intents, Topics, Citation Share and period comparison.
That means publishers can start seeing not only that they were cited, but the broader type of query, thematic areas and their proportion of citations within a given grounding query.
And in October 2026, Microsoft launched Topics Portfolio in beta, giving eligible sites a broader view of the thematic areas in which they appear, approximate topic citation demand, how many sites are cited and opportunity queries where other sites appear but theirs does not.
These are genuine advances in measurement.
They still do not tell you that a citation generated a customer.
Google Search Console has changed too
This is worth stressing because older GEO guides may now be out of date.
Google launched dedicated Generative AI performance reports in Search Console in June 2026 and says that by 31st August the insights had rolled out worldwide.
The Search report covers AI Overviews and AI Mode.
It shows generative-AI impressions and allows businesses to analyse them by page, country, date and device.
That is considerably better than having all AI visibility buried indistinguishably inside the standard Web report.
But the dedicated report still does not give you everything you might want.
It is primarily an impression visibility report.
It does not turn Google AI into an ordinary keyword-ranking report where you can see every exact conversational prompt and a neat position number.
That is why your measurement system still needs several sources.
Can you measure ChatGPT visibility directly?
There is not one simple publisher metric called "your ChatGPT rank".
OpenAI says Search placement is not guaranteed and that eligible public content can be surfaced when OAI-SearchBot has access.
So businesses currently tend to combine several methods:
manual testing
third-party monitoring platforms
website referral analytics
customer attribution
Third-party tools can be useful, but I would apply exactly the same scepticism I apply to SEO tools.
Understand what is actually being sampled.
How many prompts?
Which models?
Which geography?
How often?
Does "visibility" mean cited, named or recommended?
A beautiful dashboard can still be measuring something that has very little connection to your commercial goal.
What does Fair PR monitor in practice?
Fair is an unusually useful example because our current website is new.
We launched the current Fair PR business and website in July 2026 and began much more structured SEO/GEO work around the middle of that month.
So at the time of writing, we are looking at roughly three months of concentrated work rather than pretending we have a ten-year AI-search dataset.
We maintain a dedicated SEO & GEO Query Monitoring board.
As of my latest check, it contains 119 tracked prompts and search queries.
The live board currently records:
21 ChatGPT prompts where Fair is marked Recommended
7 where Fair is Mentioned
34 marked Not checked
57 where the current ChatGPT field remains blank
On the Google AI side, the live board currently records:
27 Included
10 Not included
22 Not checked
60 currently blank
Those blank values are crucial.
They do not mean Fair failed to appear.
They mean we do not currently have a populated observation for that row.
That distinction is exactly the sort of thing a reporting dashboard needs to make clear.
If I conveniently divided 21 recommendations by 119 prompts and presented the result as "Fair's ChatGPT market share", I would be creating a statistic that the underlying data does not support.
What do we actually record for each prompt?
We try to make the monitoring useful rather than simply collecting screenshots.
Each query can have:
a market
search intent
commercial cluster
target Fair URL
Google Search Console performance
Google AI status
Google-selected Fair page
ChatGPT status
ChatGPT-selected Fair page where identifiable
previous AI states
last-check dates
directional GEO trend
That helps us ask better questions.
Is Fair being recommended?
If not, is it at least understood?
Which page is being selected?
Is that the page we intended?
Are weaker sectors improving?
Are we strong in the UK but weak in the US for the same concept?
That becomes actionable.
One of the most useful distinctions is retrieval versus understanding
We have observed queries where Fair does not initially appear in a shortlist, but once the system investigates Fair directly it understands the company extremely well.
It can identify the journalist-led model.
Pricing.
Markets.
Services.
Earned-media positioning.
Relevant specialisms.
That tells us something quite specific.
The problem may not be that the system misunderstands Fair.
The problem may be candidate retrieval or shortlisting.
That leads to a very different website strategy than rewriting the entire site because we assume the entity is fundamentally confused.
This is why measurement should influence decisions.
How can Google Analytics help?
Traffic is another layer.
In May 2026 Google Analytics introduced a dedicated AI Assistant channel.
Recognised referrals from AI assistants can now be classified with an `ai-assistant` medium and grouped into the AI Assistant channel, giving businesses a much cleaner way to compare this traffic with traditional channels. Google specifically mentions assistants such as ChatGPT, Gemini and Claude.
That is extremely useful.
But it still only records people who arrive at your website in a way Analytics can observe.
It does not tell you how many people saw your name in an AI answer and never clicked.
ChatGPT referrals have another useful tracking signal
OpenAI says ChatGPT Search automatically adds `utm_source=chatgpt.com` to referral URLs.
That means website owners can use analytics tools to identify click-through traffic arriving from ChatGPT Search.
Again, that is referral traffic.
It is not the same as every time ChatGPT mentioned your company.
A person could see Fair recommended, remember the name, open Google later and type "Fair Public Relations".
Analytics might then classify the visit differently.
Another person could ask ChatGPT for our telephone number and never visit the website at all.
Attribution leaks.
That is normal.
What has Fair seen in GA4?
Our latest settled 28-day internal reporting period currently recorded in the website master covers 7th September to 4th October 2026.
During that period, GA4 recorded:
857 total sessions
107 sessions classified within the AI Assistant channel
7 key events from that channel
Looking specifically at recognised ChatGPT traffic across our AI/referral classification, we recorded 109 sessions, representing approximately 12.72% of total sessions, with around 57.8% engagement and 8 key events.
Those two ChatGPT/AI Assistant figures should not simply be added together.
They are different ways of looking at overlapping traffic classifications.
And, importantly, a GA4 key event is not automatically a qualified sales lead.
Our own reporting rules deliberately separate raw web analytics from actual enquiry quality.
That is how I think businesses should approach this generally.
What about the enquiries themselves?
This is where AI search has become commercially interesting for us.
We are now regularly seeing prospective clients tell us they found Fair through ChatGPT, sometimes several in a day.
That self-reported information is valuable because it closes part of the gap between an AI appearance and an actual buyer.
But even here I would remain cautious.
People do not always remember accurately where they first encountered a business.
Somebody might say "ChatGPT" because that was the last thing they used after already seeing the company elsewhere.
Self-reporting is evidence.
It is not perfect attribution.
Add "How did you hear about us?" to your enquiry process
This remains one of the simplest things a business can do.
Ask the customer.
Include options that reflect the way people genuinely discover you.
Google.
ChatGPT.
Other AI assistant.
Social media.
Recommendation.
Press article.
Other.
Then look at that information alongside Analytics and CRM data.
If twenty strong prospects say ChatGPT brought them to you, that is commercially important even if your analytics platform failed to label every journey perfectly.
What should you measure after the enquiry?
Do not stop at "AI lead".
Was the enquiry actually appropriate?
Could they afford the service?
Were they in your market?
Did they book a call?
Did you issue a proposal?
Did they sign?
How much revenue resulted?
A hundred AI-originating enquiries from people who cannot buy your service may be less useful than three highly qualified enquiries matching your ideal customer perfectly.
This is where measurement becomes commercial rather than performative.
What do I personally care about most?
For visibility itself, two things stand out to me.
The first is being genuinely recommended for a relevant buying question.
That is powerful because the system is not simply finding one of our articles. It is considering Fair as a possible answer to somebody's commercial need.
The second is being understood accurately.
If ChatGPT understands that Fair is journalist-led, founder-led, focused on earned media, works in the UK, US and Canada, has transparent pricing and serves particular types of businesses, that matters enormously.
I would rather be accurately understood and recommended for ten relevant prompts than repeatedly appear in fifty irrelevant answers.
Citations come after that for me.
Useful, interesting, important to track, but not identical to buyer visibility.
Referral traffic then tells us whether people are moving from those systems onto the website.
And ultimately, of course, the most commercially valuable metric is whether appropriate customers actually arrive and buy.
That creates two hierarchies:
visibility quality: recommendation and accurate understanding first;
business value: suitable enquiries and customers first.
They are related, but they are not interchangeable.
A practical AI visibility scorecard
| Measurement | Source | What it genuinely tells you | What it does not prove |
|---|---|---|---|
| AI citations | Platform reporting such as Bing AI Performance or Google Generative AI report | Your content is being surfaced as supporting material within the platform's measured AI experiences | That the brand was recommended or a customer saw it |
| Brand mentions | Repeated prompt observations or monitoring software | The company name appears in answers | That it is consistently recommended |
| Relevant recommendations | Controlled buyer-prompt monitoring | The business is being suggested for commercially relevant questions | Guaranteed visibility for every user |
| Description accuracy | Human review | Whether the AI understands the business correctly | Commercial intent |
| AI referral traffic | GA4 / referral data | Observable click-through visits | Total exposure |
| Self-reported AI enquiries | Forms / CRM / calls | Customers believe AI influenced discovery | Perfect first-touch attribution |
| Qualified enquiries | CRM / sales records | AI-associated prospects fit the actual offer | Revenue until they convert |
| Customers and revenue | Business records | Commercial value where attribution is credible | Every wider reputation benefit |
I think this is much more useful than one "AI authority score".
How often should businesses check AI visibility?
Not every five minutes.
Daily checking may be useful for somebody developing a specialist monitoring platform or studying a fast-moving launch.
For a normal small business, it can create noise.
A weekly or monthly structured review is often much more useful, depending on how many prompts and platforms are being monitored.
The important thing is to maintain enough consistency to notice patterns.
You could combine:
periodic prompt observations
continuous first-party platform data
monthly GA4 analysis
ongoing enquiry attribution
That gives you several independent evidence streams.
Look for trends, not dopamine
There is something psychologically satisfying about repeatedly asking ChatGPT questions until it finally says your company's name.
That is not measurement.
If you test twenty slightly different versions until one produces the answer you wanted, screenshot that one and put it in a client report, you have demonstrated persistence rather than visibility.
A good monitoring process should retain the misses too.
Fair's board records Not included and Not checked alongside Recommended and Mentioned.
That makes the data less flattering.
It also makes it more useful.
What should an honest GEO or PR report contain?
I would want to know what the agency actually did.
What was published?
What changed?
What is directly measured?
Which observations are manual?
What platform generated the data?
What remains uncertain?
If a brand appeared in ChatGPT, what was the question?
Was it recommended or simply mentioned?
Did a relevant page get cited?
Was the answer accurate?
Did traffic arrive?
Did an enquiry result?
And what are we going to do differently because of that information?
The report should help make the next decision.
It should not simply create a colourful document proving the agency deserves another month's retainer.
Be extremely careful with before-and-after causation
Suppose a PR agency secures a newspaper article on Monday.
On Friday, the company notices that ChatGPT recommends it.
Can the agency say:
"Our article caused the ChatGPT recommendation"?
Usually, no.
The recommendation may have changed because of many things.
A model or retrieval-system update.
Different prompt wording.
A website update.
Another source.
Fresh search results.
Location.
Changing competition.
Or the press article.
The article may have contributed.
That is a hypothesis unless you have stronger evidence.
Responsible reporting should say exactly that.
Does more AI visibility always mean more business?
No.
The same problem exists in conventional SEO.
Traffic is not automatically commercial value.
Visibility is not automatically demand.
A company can become extremely visible around informational questions answered mostly by students and journalists while remaining invisible when buyers are actually choosing a supplier.
That may still have authority value.
But it is not the same outcome.
This is why I keep returning to real buyer prompts.
If we want Fair to grow, I care whether somebody asking for a PR agency that fits our actual offer sees us.
I care much less whether ChatGPT cites our article in an academic explanation of the history of public relations.
What can Fair's first three months actually prove?
Not much conclusively.
And that is part of why I think the example is useful.
We can say the website is receiving substantial AI referral traffic.
We can say customers are reporting ChatGPT as a discovery source.
We can say the company is being recommended for a growing number of monitored commercial prompts.
We can say Google's generative search is surfacing Fair for multiple areas.
We can say the site is still relatively new.
What we cannot say is:
"We followed these twelve steps and therefore increased ChatGPT rankings by 300%."
There was no universally comparable starting "ChatGPT ranking".
Our website work also includes SEO, content, commercial pages, structured information, case studies, internal linking and wider positioning.
PR exists around the business.
Search systems change.
We therefore have a very promising body of evidence.
Not a controlled scientific experiment.
That is the level of confidence I want Fair to maintain in its own marketing too.
Frequently asked questions
Can you check your ChatGPT ranking?
You can monitor how your business appears for a defined set of prompts, but there is not one universal public ChatGPT ranking comparable to a simple search-engine keyword position.
Does ChatGPT always recommend the same businesses?
No. Answers and sources can change depending on the question, context, available information and changes to the system.
How do you know whether ChatGPT cited your website?
When ChatGPT Search provides citations, you can inspect the cited sources. Website analytics can also identify some resulting referrals because OpenAI adds `utm_source=chatgpt.com` to ChatGPT Search referral links.
Does Google Search Console show AI Overview visibility separately?
Yes, now. Google launched dedicated Generative AI performance reporting in 2026 covering AI Overviews and AI Mode and says it had rolled out worldwide by 31st August.
Can GA4 identify ChatGPT visitors?
Some of them. GA4 introduced an AI Assistant channel in May 2026 for recognised AI-assistant referrals, and ChatGPT also adds a trackable UTM parameter to its Search referral links. Neither captures every possible AI-influenced customer journey.
Microsoft's Bing Webmaster Tools defines Citation Share as the percentage of citations attributed to your site out of all citations shown across sites for the same grounding query. Microsoft describes it as observational, not as a ranking system.
How often should you test AI recommendations?
There is no scientifically universal interval. Choose a repeatable schedule that suits the business and compare like with like. For many small businesses, regular weekly or monthly sampling is more practical than obsessive daily checks.
Is an AI brand mention valuable if nobody clicks?
Potentially. It can create awareness and influence later behaviour, but the commercial value is more difficult to attribute without a click or customer self-report.
Can PR agencies prove that media coverage caused an AI recommendation?
Usually not. An agency can prove the coverage exists. It may observe AI changes afterwards. Direct causation requires much stronger evidence than timing alone.
What should I ask an agency to include in a GEO report?
Ask for clear separation between actual citations, mentions, recommendations, referral traffic, enquiries and sales. Ask what is measured directly, what is sampled manually, what remains uncertain and how the findings are changing the strategy.
Want PR that builds genuine credibility, not just impressive-looking numbers?
At Fair Public Relations, our core job is still PR.
We develop positioning, stories and expert commentary, approach journalists and build earned editorial coverage around things our clients can genuinely substantiate.
We are increasingly interested in how that wider public profile interacts with AI discovery because we are seeing it happen in our own business.
But we would rather show you what we can actually measure than invent a guaranteed GEO result.
If you want to explore whether Standard PR could help strengthen the credible public information around your business, you can get in touch and we will tell you whether we think there is genuine media potential.
About the author
Jordan Baker is the founder of Fair Public Relations and a former UK national journalist with more than a decade of experience in journalism, PR and media relations. He writes about earned media, AI search, GEO measurement, journalist outreach and how small businesses can build measurable visibility without pretending every result can be perfectly attributed.
View Jordan Baker's media expert profile
