What we did

On 27 and 28 September 2026 we collected 496 Perplexity answers to questions in the form "best in , UK". There were 456 different questions (some we asked twice) across 105 professions: dentists, accountants, architects, physios, estate agents and so on. We took the links Perplexity shows inline in each answer and counted which sites they point to.

One limit up front. This is basic mode, and we only counted the inline links, not the full source panel. Numbers below describe this sample, not everything anyone sees.

The short version

A typical answer named 5 firms (between 0 and 14) and cited a median of 15 sources. Only 3 of the 496 answers named no firm at all, so Perplexity almost always gives you a list.

Now the part that surprised us.

• Google Maps or Google Search pages as a source: 0 of 496.

• A Google Business page: 1 of 496.

A listing many local firms work hard on almost never shows up in the linked sources.

What does? Directories and review sites. At least one directory or review site (Yell, Bark, Trustpilot and similar) appeared in 410 of the 496 answers. The most-cited domains were trustpilot.com in 190 answers, yell.com in 87, yelp.com in 82, whatclinic.com in 60, checkatrade.com in 59, and doctify.com, fresha.com and trustatrader.com in 54 each. But presence is not dominance: directories and review sites made up more than half of the sources in only 8 of the 496 answers.

A few smaller findings are worth a line. Professional registers and regulators (ICAEW, SRA, GDC, RICS, CQC, gov.uk and similar) were cited in 67 answers. Reddit was cited in 3, Facebook or Instagram in 38.

Some trades have one favourite site

This is the part we would act on first. When we split by profession, in these five trades one site appeared in more than half of the answers:

• Estate agents: getagent.co.uk in 24 of 26 answers.

• Architects: houzz.co.uk in 23 of 25.

• Accountants: businessfinancing.co.uk in 26 of 48.

• Physiotherapists: fresha.com in 31 of 55.

• Dentists: bestdentists.uk and mynextdentist.co.uk in 21 of 31 each.

So "be visible to AI" is not one job. For an estate agent, a good first question is "what does getagent.co.uk say about my branch". For an architect, "when did I last touch my Houzz profile".

We also looked at "best X in town" list pages. Of the 250 answers where we had full source URLs, 52 cited one of those third-party lists. If you are not on them, the answer is unlikely to find you through them.

Agencies are in the same boat

We ran "best SEO agency in " for 30 UK cities on 28 September, one run per city. Clutch was cited in 17 of the 30 answers, The Manifest in 11 and DesignRush in 9. On 25 September, a run with two answers per city had Clutch in 33 of 60. Same pattern: directories sit inside a lot of these answers, even for SEO agencies themselves.

The answer changes when you ask twice

On 28 September we asked 40 questions twice, about three hours apart. The two answers named the identical list of firms in only 3 of the 40 pairs. In one pair they had no firm in common. On average the two lists shared 48% of names, and of the 324 firms named across those pairs, 179 showed up in only one of the two answers.

This matters for how you check yourself. One look at Perplexity tells you very little. Ask the same question a few times, on different days, and note how often you appear. If you see yourself once, that is not a result.

What we would do with this

None of this is a promise that a fix will put you in the answer. We can't see inside the model, and the answer moves from run to run. But the sources are visible, and that part you can check in ten minutes:

1. Search "best in " in Perplexity.

2. Click every link in the answer. Write down the domains.

3. Check which of them list you, whether the details match your own site, and whether the profile is current.

4. If your trade has a register or regulator, make sure your entry matches your site.

5. Repeat on another day and compare.

You will quickly see which two or three sites keep coming up for your trade. That is where the effort goes, not into a generic checklist.

Sources and method

The full write-up with the source counts by profession is on our site: Which sites Perplexity cites for UK local firms. The way we measure, and why we run each question more than once, is here: How to measure AI search visibility.

CitedGEO is a GEO agency. We measure how often brands are named and cited in Perplexity and Google AI Overviews, and publish how.

That's not a design opinion. It's a field check.

Here's what the two missing fields look like, and how to check yours in a minute.

What "before" and "after" looks like

Here's a generic example, not a real shop. The product is invented.

Invented example. Left: the offer block most shops had. Right: the two blocks that were missing on 35 of the 37 pages we opened.

What we actually checked

We opened one product page per shop, the first one we could reach from the homepage. 37 of 80 opened as a real product page, with a price or a basket button. The other 43 didn't: we couldn't find a product link from the homepage, robots.txt or the page blocked us, or the page was a catalogue with no price and no basket. We're not counting those 43 either way. This is a report on the 37 pages we could read.

On each of the 37, we looked at the schema.org Product and Offer markup, the JSON-LD block most shops already have for search results and price snippets. We checked for four things: is there a Product entry at all, does it have an offer, does that offer carry a price and stock status, and does it carry shipping details and a return policy.

33 of 37 pages have a gap. 30 of those are missing only shipping and returns, the two fields Google specifically asks for.

30 of the 37 have Product and Offer markup with a price and stock status, but nothing about shipping or returns. 2 have no Product markup at all. 1 has an offer with no price or availability. 4 passed our check. But only 2 of those actually carry shipping or return fields; the other 2 use older inline markup that our check doesn't read field by field, and the two fields aren't on those pages either.

We also searched each whole product page and each homepage, not just the Product block, for a policy set once at the organisation level. None of the 80 shops had one there.

Why these two fields specifically

Google's own documentation on merchant listing structured data lists shippingDetails and hasMerchantReturnPolicy as recommended properties on a product Offer, the fields that let a listing show delivery and return information in Search (developers.google.com/search/docs/appearance/structured-data/merchant-listing). Google suggests setting them once at the organisation level and only overriding per product when a shop's policy differs. Without either, a listing just skips that information.

We're not claiming an AI shopping assistant definitely reads these two fields the way Google's merchant listings do. We checked one product page per shop, not the whole catalogue, and we checked on one day, 30 September 2026. What we can say is narrower and still useful: these are the fields Google names for its own merchant features, and 35 of 37 shops don't have them, regardless of what any AI tool does with that gap.

Where the 80 shops came from

We started from Companies House: active limited companies with SIC code 47910, "retail sale via mail order houses or via internet". We kept only the ones where the live site itself names the company, by Companies House number or by name. 80 shops passed that bar.

On the homepage, we could tell the platform for all 80: Shopify on 38, WooCommerce on 18, something else on the remaining 24.

Shopify and WooCommerce cover 70% of the 80 shops we found.

Check your own page in the next minute

Open Google's Rich Results Test, paste in one of your product page URLs, and look at the Product result. Or just view the page source and search for "shippingDetails" and "hasMerchantReturnPolicy". If neither string appears, you're with the 35.

Of the pages we could read, 28 were Shopify and 8 WooCommerce; the gap showed up on both. If you're on Shopify, this markup usually comes from your theme or a structured data app. Look there first rather than editing code by hand. If you're on WooCommerce, the same applies: it's your theme or a schema/SEO plugin that should be emitting this block. Check its settings before writing JSON-LD from scratch. Either way, the fix is adding two fields to markup you very likely already have, not building a new system.

What we're not saying

We're not saying add this markup and AI search will start naming you. We're saying Google has told store owners for years what it wants to see on a product page, and two years into "AI search changes everything," most of the shops we opened still haven't done the first, smallest thing Google asked for.

CitedGEO checks what your product pages actually say to search engines and AI tools, and what's missing.

The fixes described here have been made. This article reports the baseline measurement on 13 September 2026, what the audit found and what was changed. Re-measurement is scheduled; results will be published.

The business

• An online retailer of professional kitchen equipment.
• A built-to-order catalog of roughly 5,400 active products.
• No warehouse stock.

A catalog like this has two traits that matter for AI search. It is large, so small template problems repeat thousands of times. And it competes in categories where well-known equipment brands already have strong visibility of their own.

Starting point: 1 of 15 queries

On September 13, 2026, we measured the client's presence in AI answers with 15 commercial search queries run through Perplexity, covering the main equipment categories it sells.

The result:

• The client's site was cited in 1 of the 15 queries, for a narrow product category.
• In that one answer it ranked third of seven sources, with three citations inside the answer text.
• The other 14 queries did not mention the client.
• One competitor was cited in 9 of the 15 queries, well ahead of the field.
• Four other competitors were each cited in 3 to 4 queries.

A note on scope: this baseline was a single engine, Perplexity, and 15 queries. It should be read as a first look at one engine, not as a full picture: our audits today run Perplexity and Google AI Overviews.

The explanation that did not hold up

The obvious explanation for 1 of 15 was that the cited page must be the strong one and the rest were thin. The usual advice follows: add more content to the other category pages.

We tested that directly. We compared the one cited category page against a major category that was never cited:

On every measure the uncited category was the stronger page. It had more body text, an FAQ block the cited page lacked, more linked articles and more products. Content length, an FAQ block and cluster size did not explain the gap.

The likely remaining factor, which we have not separately verified: the uncited category has established brand-name competitors already occupying AI answers; the cited one does not.

This is worth dwelling on. "Write more" is a common recommendation in AI visibility advice, and here the page with 51,036 characters of body text was the one that was never cited, while the page with 17,946 was the one that got cited. The finding does not prove that competition is the cause. It does show that the content-volume explanation was wrong for this site, which changes where the effort should go.

What Search Console showed

A separate pull of three months of Search Console data on September 15 added a second angle:

• Articles drove 42% of the site's total clicks.
• Product pages, which make up 74% of indexed URLs, drove 11% of clicks.
• Local-language product pages had the best position (9.8) and CTR (2.2%) on the site.

The last point narrowed the problem. Those pages were reaching the index; they were not getting shown. For a retailer, that is a different problem from "Google cannot find the products", and it needs different fixes.

What GA4 showed about AI referrals

A GA4 pull the same month looked at the AI-assistant referral channel: visitors arriving from AI tools.

• Over three months, July through September, the channel produced zero recorded conversions.
• Over the same period, engagement on that channel rose:
  • session duration went from 16 to 34 seconds;
  • events per session went from 5.17 to 7.63.

Traffic and engagement from AI referrals were growing; conversions were not. The next section also records a tracking problem that limits how far that conversion number can be read.

What we did

The work fell into four groups: access for crawlers, data the engines and shopping surfaces read, crawlability of the catalog, and measurement.

Crawler access

• llms.txt. Published an llms.txt file listing the 20 top-level categories with product counts and 19 brands.
• Firewall. Corrected a firewall rule that was blocking Meta's link-preview crawler, and confirmed that the major AI crawlers (GPTBot, OAI-SearchBot, PerplexityBot) and Googlebot were reaching the site unblocked.

Structured data and the product feed

• Product schema. 4,078 of 5,422 products had an empty description field in their schema markup, because the fallback source field was never wired in. We added it, and matched spec attributes to the visible specification tables.
• Google product feed. The static feed file was frozen behind a 10-year edge cache. We rebuilt the feed as a dynamic endpoint; it now serves 4,577 products per language on each request.

The schema fix is a good example of a problem that is invisible from the front end. The product pages looked complete to a human visitor. The structured data behind most of them had no description at all, which is the part of the page written for machines rather than people.

Catalog crawlability

• Pagination. Category pagination pages were marked noindex, nofollow, which had made listing pages past page 1 invisible to crawlers. Fixed.
• Duplicate branch. Removed a duplicate category branch that was splitting the same products across two listings.
• Internal search. Fixed internal search so that different dash characters in size specs (for example "1500x600" versus "1500×600") return the same results.

Measurement

• Analytics. Fixed duplicate purchase-event tracking so that conversion data, including from AI referrals, can be measured going forward.

This last fix matters for how the GA4 numbers above should be read. With duplicate purchase events, conversion data could not be relied on. The fix does not change the past three months; it is there so the next measurement can be read properly.

Status

Re-measurement is scheduled; results will be published. It will run the same 15 queries through Perplexity and compare them with the baseline of 13 September, when the retailer was cited in 1 of 15. We are not claiming that citations increased, that the leading competitor was overtaken, or that AI referrals started converting.

What other retailers can take from this

Without adding anything the case does not show, a few practical points carry over to other large catalogs:

1. Test the obvious explanation before acting on it. Comparing a cited page with an uncited one ruled out the most common recommendation before any time was spent on it.
2. Check structured data at catalog scale. A template gap in one field can affect thousands of products while every page looks fine in a browser.
3. Check what the firewall and cache are doing. A firewall rule and a long edge cache were both quietly shaping what crawlers and feeds received.
4. Look at pagination and duplicates. A noindex, nofollow on pagination makes listing pages past page 1 invisible to crawlers.
5. Fix tracking before judging a channel. Zero recorded conversions from AI referrals means little if purchase events are duplicated or broken.

About CitedGEO

CitedGEO is a Generative Engine Optimization agency; this case study and our measurement method are published on our site. If you want a baseline for your own category, the free GEO audit runs 10 of your buyers' questions through Perplexity and Google AI Overviews, 3 times each, and shows where you appear, where competitors appear, and what to fix first. It is a measurement and a ranked list of fixes, with no promise about what any engine will say afterwards.

This article walks through the measurement method we publish on our site, piece by piece: what we count, how we build the list of questions, how often we run it, how we keep one month comparable with the next, what happens when a model changes underneath us, and where the method stops being reliable. Every rule described here is taken from the public method page at citedgeo.com/measurement/. We explain the reasoning around each rule, but we do not add numbers the method does not contain.

What "AI visibility" means

We define AI visibility as how often, how prominently and how accurately a brand appears in answers that AI engines give to the questions its buyers ask.

Frequency needs a fixed set of questions and a fixed number of runs, or it cannot be compared over time. Prominence is about placement: being one of several brands in a list is not the same as being the one the answer recommends. Accuracy is about what the engine says, because a mention that gets your category or location wrong can do more harm than no mention.

The six metrics and their formulas

We track six metrics. Each one has a written definition and a formula, so anyone reading a report can recompute it from the raw answers.

A few notes on how these relate to each other:

Mention and citation are different events. An engine can name a brand in the text without linking to its site, and it can list a domain as a source without naming the brand in the answer. Engines that search the web live, such as Perplexity, usually show their sources; that is where citation rate becomes a useful signal of whether your own pages are being used as evidence.

Recommendation is stricter than mention. An answer that says "brand X also exists" counts as a mention. An answer that presents brand X as an option to choose counts as a recommendation. Buyers act on the second.

Share of voice needs a competitor list. The denominator is mentions of you plus mentions of the competitors you agreed to track. Change the list and the number changes, which is why the competitor set is fixed for a reporting period.

Accuracy uses mentions as its base, not answers. It asks: when the engine talks about you, is it right?

Why we never blend engines into one score

We report every metric per engine and per market. We do not combine Perplexity and Google AI Overviews into a single "AI visibility score", because a blend hides where the problem is. A brand can be well covered in one engine and absent in the other; an average of the two looks moderate and points to nothing.

Why 10 prompts × 2 engines × 3 runs

The free audit has a fixed default size:

• 10 prompts: discovery, comparison, problem and brand questions for the category.
• 2 engines: Perplexity and Google AI Overviews.
• 3 runs of each prompt on each engine.

That gives 10 × 2 × 3 = 60 recorded answers, collected once. The paid GEO audit uses the same structure at a larger size: 30 buyer questions, run 3 times each in Perplexity and Google AI Overviews, for 180 recorded answers.

The third number is the one people question most often. Why ask the same thing three times?

Because AI answers are not deterministic. The same prompt can name different brands on different runs. If you run a prompt once and your brand appears, you do not know whether it would appear the next time. A rate over 3 runs is steadier than a single observation, and for prompts that turn out to be volatile we can raise the number of runs.

How we build the prompt set

The prompt set is the foundation of every number, so it is built before the first run and agreed with the client.

Where the questions come from

• Your sales and support questions: what prospects actually ask before they buy.
• Your search console queries: the language people already use to find you.
• Reviews: what customers say mattered to them.
• Community threads in your category: how buyers describe the problem in their own words.

Which intents it covers

The set mixes five kinds of prompts: category discovery, comparison, problem, brand and local prompts. A discovery prompt asks "what are the options for X"; a comparison prompt puts two or more brands side by side; a problem prompt describes a situation without naming a product; a brand prompt asks about you directly; a local prompt adds a place.

How it stays stable

• A fixed core set that never changes within a reporting period, so trends are comparable.
• A smaller exploratory set that rotates, to catch new questions buyers start asking.
• Versioning: every change to the set is logged with a date and a reason.

How often we run it

How we keep runs comparable

A trend line only means something if each point on it was produced the same way. Four rules keep runs comparable:

1. Clean sessions with no personal history, where the platform allows it.
2. A fixed language and region per market.
3. Metadata with every answer: the date, platform, surface and model name (where the platform shows it) are stored alongside the answer text.
4. The same classification rules every time, with a sample checked by a second person.

Deciding whether an answer "recommends" a brand or merely "mentions" it is a judgement; written rules and a second reader keep it from drifting.

APIs versus consumer apps

We record which surface each answer came from. Consumer apps and API requests can answer differently, so we do not mix them in one metric.

What happens when a model is updated

Model updates are the main threat to any AI visibility trend. A jump or a drop after an update may have nothing to do with anything the brand did. Our procedure has five steps:

1. Detect. From the vendor's release notes, or from a sudden shift across the core set.
2. Re-run. Run the full prompt set on the new model.
3. Mark the break. The trend chart shows a break, and we do not compare numbers across it without saying so.
4. Re-map sources. Check whether the new model cites different domains for your topics.
5. Adjust the plan. Change next month's priorities if the source map changed.

The re-mapping step is often the most useful. If a new model starts citing a different set of review sites or directories for your category, the work of getting your brand described correctly on third-party sites needs to follow it.

Limitations

A method is only as honest as its stated limits. Ours has three:

• Answers vary between runs and between users. Our rates describe a sample, not every answer anyone sees.
• Personalisation and logged-in history can change answers in ways a clean session does not show.
• We cannot see inside the models. When a number moves, we separate what we can show from what we suspect.

There is also a timing limit worth knowing. Engines that search the web live can reflect changes once your pages and third-party mentions are recrawled. What a model knows from training changes only when the model is retrained. That is why we record a baseline first and report the change every month, including when nothing has moved.

Where measurement sits in the Citation Loop

Measurement is not a separate product; it is the first and last step of a monthly cycle we name the Citation Loop:

1. Probe: run your buyers' questions through Perplexity and Google AI Overviews, 3 times each, and record who is named and who is cited.
2. Diagnose: list the sources the engines cite for your topic and compare what they say about you with what is true.
3. Build: fix crawl access and page structure; publish definitions, comparisons and facts an engine can quote.
4. Earn: get your brand described correctly on the third-party sites the engines already cite.
5. Re-measure: run the same prompt set again, compare with the baseline and pick the next cycle's priorities.

Each cycle ends with a measurement, and that measurement sets the next cycle's work.

Measurement FAQ

Can I see the raw answers? Yes. Every report includes the prompt list and the raw answers behind each number.

What if our numbers go down? We report it, with the likely reasons we can support with data, and say plainly where we do not know.

Can anyone promise that Perplexity or Google AI Overviews will mention my brand? No. No agency controls what a model outputs. What can be done is to improve the signals AI engines rely on and measure the change with a published method, so you can judge the result yourself.

About CitedGEO

CitedGEO is a Generative Engine Optimization agency that publishes how it measures AI visibility, including the method described above. If you want to see where your brand stands today, the free GEO audit runs 10 buyer prompts through Perplexity and Google AI Overviews, 3 times each, and shows who is named and cited for your category, along with a ranked list of what to fix first. It is a baseline measurement, not a promise of any particular outcome.