On 29 July 2026, Search Engine Land published analysis showing that roughly 40% of AI citations, across about 16 million brand appearances, never name the source brand in the answer text. One day later, on 30 July 2026, the same publication covered a geoSurge study finding that AI models search for brands they already recognise around 3.2 times more often than brands they do not. Read together, those two findings describe a specific problem for a small firm: your writing can win the answer while your name loses it.
What actually changed in AI search this month?
The ghost citation analysis reported by Search Engine Land on 29 July 2026 defines the term plainly: a citation does not mean your brand was seen. It quantifies the gap across seven engines. The geoSurge work, published the following day, explains part of why the gap exists. If a model reaches for names it already holds, the unfamiliar name attached to a genuinely useful page is the easiest thing to drop when the answer gets compressed.
For a UK owner writing their own marketing, this reframes the job. The target was never only the link.
What is a ghost citation, and why does it matter to a small business?
The scale differs by study, which is worth stating honestly rather than picking the bigger number. Search Engine Land's July figure is roughly 40% across about 16 million brand appearances. An earlier and smaller Semrush study published on 9 June 2026, authored by Margarita Loktionova with Christine Skopec and produced with Kevin Indig of Growth Memo, put the rate closer to 62% across just under 4,000 domain appearances from 115 prompts in 14 countries. That study also found only around 13% of appearances produced both a citation and a brand mention.
Different samples, different prompt sets, same direction of travel.
Which AI engines are worst for ghost citations?
The article groups engines into namers and citers. Gemini and Copilot mention brands more often but cite less frequently. Perplexity and Google link generously and drop names more often. That distinction matters because it changes where effort pays.
A UK business whose customers overwhelmingly start in Google is exposed to the citer pattern by default. The Semrush data adds a sharper edge: ChatGPT showed a citation rate near 87% against a mention rate around 21%, while Gemini ran roughly the inverse, citing about 21% of the time and naming brands in over 80% of answers. The study's own summary notes there is almost no overlap between which brands ChatGPT cites and which Gemini names for the same prompt.
Two engines, same question, two entirely different visible winners.
Why do AI models keep picking brands they already know?
Two further numbers from that study are worth an owner's attention. Only about 31% of fan-out searches named any brand at all, meaning roughly two thirds were generic category searches. And where a brand was named, about 63% of those searches targeted one of the model's top five recalled brands.
The researchers state clearly that this measures an association in exploratory data, not a proven cause. They also note the escape route. Live search can still find unfamiliar brands, particularly in categories where the model relies less on stored knowledge. Local trades, regional specialists and niche services sit squarely in that territory, which is why the useful response is entity clarity rather than despair. The distinction between optimising for classic rankings and optimising for machine answers is explored further in this breakdown of AEO, GEO and LLMO for UK firms.
Is AI search really taking clicks away from UK small business websites?
The Search Engine Journal write-up of that revised working paper reports that AI Overviews triggered on roughly 41% of all queries and 53% of informational ones. It also dismantles a common defence. Google's Liz Reid has argued that summaries mainly remove low-quality bounce clicks. The researchers reported no statistically significant difference in bounce rate, time on site or return-to-search between the two conditions, with bounce around 40% and ten-second exits around 18% in both groups.
When Overviews were removed, the top three results gained most of the recovered clicks, with position one close to doubling.
Which UK businesses should worry least about this?
That is not a reason for a plumber or a dental practice to ignore the shift. It is a reason to sequence the work correctly. Fixing the transactional path, meaning the profile, the local pages, the booking route, still returns faster than rewriting a blog archive.
What should a UK owner actually change on their website?
The first is the cheapest and most overlooked. Advice written in an anonymous voice invites anonymisation. A sentence that reads "most Leeds installers quote between £180 and £240 for this" survives the summariser with no name attached. A sentence that reads "Northgate Heating quotes between £180 and £240 for this in Leeds" is harder to strip without breaking the fact. Search Engine Land's recommendations put this plainly: pair attribution closely, placing brand names directly with claims rather than distantly.
The second is proprietary research. A named study, a regional pricing survey, a documented method, these are harder to anonymise because the name is part of the fact.
The third is unglamorous entity hygiene, and it is where most UK small businesses actually lose. Three variants of a trading name across a website, a Google Business Profile and a directory listing gives a model three weak entities instead of one recognisable one.
How do you get named, not just cited?
Practical starting points that need no new tooling:
Rewrite the top ten commercial pages so the trading name appears inside claim sentences, not only in headings. Add a short, dated methodology note to anything containing original figures. Publish prices or price bands where the business is willing, since specific numbers attached to a named source are among the hardest things for a summariser to launder. Make sure the About page states plainly what the business does, where, and since when, in sentences a model can lift as fact.
There is a longer treatment of the prompt-testing side of this in this guide to appearing in ChatGPT answers as a UK business.
How should this be measured without expensive tools?
Search Engine Land's July 2026 recommendations point the same way: track mentions and citations as distinct metrics, and analyse individual answers rather than only aggregate scores, because reading a specific answer is what reveals why the name was dropped.
Log competitor names that appear beside yours. If the same three names recur across engines, that is the familiarity effect from the geoSurge data showing up in your own category, and it tells you which brands the models already hold.
What is the realistic trade-off here?
The honest position is that nobody can promise a named mention. The geoSurge researchers themselves flag their finding as association rather than causation. What the combined July 2026 evidence supports is narrower and still useful: citations are not mentions, mentions are not clicks, and a business measuring only one of the three is flying on a single instrument.
The businesses least affected are the ones whose customers already search the name. Which is, in the end, the same conclusion arrived at from a new direction. Build the name and the machines have something to keep.
A structured approach to tracking whether AI engines name a business, and what to change when they do not, is covered under AI visibility.
Sources accessed 31 July 2026: Search Engine Land, ghost citations, 29 July 2026; Search Engine Land, AI models favour familiar brands, 30 July 2026; Semrush and Growth Memo ghost citations study, 9 June 2026; Search Engine Journal, AI Overviews lost clicks study, 1 July 2026.
Ghost Citations — FAQ
What is a ghost citation in AI search?
A ghost citation is when an AI answer engine pulls information from your page and links to it as a source, but never writes or says your business name anywhere in the answer itself. The reader gets the advice you wrote, sees a small link icon or a source panel, and moves on without ever registering who you are. Search Engine Land, reporting on 29 July 2026, found that roughly 40% of AI citations across about 16 million brand appearances did not name the source brand. For a small business this matters because being useful to the model is not the same as being memorable to the customer. Traditional SEO measured position and clicks. AI search adds a third variable that most owners are not tracking at all: whether your name survives the summarisation step.
Which AI engines are most likely to cite without naming?
The rate varies a lot by engine. Search Engine Land's 29 July 2026 analysis put Perplexity highest at around 52%, followed by Google AI Mode at 49%, Google AI Overviews at 41% and ChatGPT at 37%. Gemini came in at 25%, Grok at 22% and Microsoft Copilot lowest at 19%. The article splits engines into two groups: namers such as Gemini and Copilot, which mention brands more but cite less often, and citers such as Perplexity and Google, which link generously but strip names from the prose. That split has a practical consequence. If most of your enquiries arrive through Google surfaces, you are more exposed to the citation-without-mention pattern than a business whose audience leans on Copilot.
Do AI models favour big brands over small UK businesses?
The evidence suggests a strong pull towards familiarity. A geoSurge study running from 29 May to 9 June 2026, covering 66 buyer prompts run 60 times each for close to 4,000 responses and over 13,000 fan-out searches, found that models searched for brands they already recognised about 3.2 times more often than brands they did not, at roughly 56% versus 17%. Where a brand-specific search happened at all, about 63% targeted one of the model's five most familiar brands. The researchers are explicit that this is an association in exploratory data, not proven causation. They also note that live search can still surface unfamiliar brands, particularly in categories where the model leans less on stored knowledge, which is exactly where local and specialist UK firms operate.
Is AI search actually costing small businesses clicks?
On informational queries, the reported answer is yes. A randomised field experiment by Saharsh Agarwal and Ananya Sen, covered by Search Engine Journal on 1 July 2026, measured a fall of around 40% in organic clicks when AI Overviews were shown, with Overviews triggering on roughly 41% of queries and 53% of informational ones. The important nuance for trades and local services is that the same study found no measurable change on navigational and transactional queries. Somebody typing an emergency callout query or a business name is still behaving much as before. The researchers also reported no statistically significant difference in bounce rate, time on site or return-to-search between clicks with and without Overviews, which undercuts the argument that the lost clicks were low quality.
What should a small business change first?
Start with name-and-claim proximity. Most pages bury the trading name in a header or footer and then write advice in an anonymous voice, which makes the brand trivially easy for a summariser to drop. Rewrite the key claims so the business name sits in the same sentence as the fact, the price band or the guarantee. Second, publish something that cannot be paraphrased without attribution: a named survey of your own customers, a regional price index, a documented method. Third, keep your name, address and phone identical everywhere, since consistency is what lets a model treat you as one entity rather than several fragments. None of this requires new software, only a rewrite pass and some discipline.
How do you measure whether this is happening to you?
Run the same ten questions a customer would ask, once a month, across ChatGPT, Gemini, Google AI Mode and Perplexity, and record two separate columns: was the site cited, and was the name written in the answer text. Those are different outcomes and collapsing them into one visibility score hides the problem. Search Engine Land's July 2026 write-up makes exactly this point, recommending that mentions and citations be tracked as distinct metrics and that individual answers be read rather than only counted. Keep the prompt wording fixed so month-to-month comparison means something, and log which competitor names appear alongside yours. Ten prompts and a spreadsheet will tell you more than most paid dashboards.
