Being cited in AI is not the same as being recommended

There is a new line item appearing on marketing invoices across the country. AI visibility tracking. Citation monitoring. Answer engine reporting. The dashboards look impressive and the numbers go up.

The problem is that the headline number on most of these reports is measuring something that does not reliably lead to customers.

The 69% problem

Research published this summer looked at what happens when a brand’s own content is cited as a source inside a Google AI Overview. You would assume that being the source of the answer means being the brand in the answer.

It does not. As the August round-up of search research sets out, in 69% of cases the brand whose page was cited was left out of the actual recommendation. Google read the page, extracted what it needed, and then recommended competitors mentioned within that page instead.

Read that again if you are paying for citation tracking. Your content did the work. Somebody else got the enquiry.

Citation counts are easy to measure, which is why so many tools lead with them. Recommendation is harder to measure and is the thing that actually drives consideration. Optimising for the easy number is optimising for the wrong one, and it is a more sophisticated version of the problem we described in can AI find your business.

Local rankings do not transfer either

The local finding is starker still, and it should give pause to anyone who has spent years and a lot of money on local SEO.

A study compared how often a set of brands appeared in Google’s local three-pack against how often AI tools recommended those same brands. The three-pack appearance rate was 35.9%. ChatGPT recommended the same businesses 1.2% of the time. Gemini managed 11%, Perplexity 7.4%.

That is roughly a thirty-fold gap for identical businesses. It implies local SEO performance and AI recommendation run on almost entirely separate signals.

This is not an argument for abandoning local SEO. Google is still where most searches happen and the three-pack still generates real enquiries. It is an argument against assuming that strong local rankings mean you are covered in AI search. They are two different battles and only one of them is being fought by most businesses.

Paid does not buy your way in

The same pattern shows up on the advertising side, which matters for anyone weighing up where to put budget.

A study of more than 50,000 commercial keywords found AI Mode showed text ads on 29% of commercial searches, rising to 54% on keywords with a cost per click above ten dollars. So the ad inventory is real and growing.

But only 11.5% of advertiser domains were also cited on the same queries where they were advertising, and just 2.3% of advertised URLs ranked organically for those terms. Paid, cited and organic are three largely non-overlapping channels inside the same interface.

Running an ad on a query does not improve your chances of being cited or recommended for it. If someone has sold you paid AI placement as a route to broader visibility, that is not what the data shows. Judge it on its own conversion merits, as you would any other paid channel, which is the same argument we made about the commercial maths in ChatGPT ads and North East B2B.

What Google now gives you, and what it withholds

On 3 June 2026 Google launched dedicated generative AI performance reports in Search Console. You will find them under Performance, in a tab alongside the standard search results view. There is a separate report for Discover.

This is genuinely useful and it closes a real gap. Until now, AI Overview and AI Mode visibility was blended into standard organic reporting with no way to isolate it.

Be clear about what it contains, though. According to Google’s own documentation, it shows impressions, pages, countries, devices and dates. It does not show clicks. It does not show click-through rate. It does not show average position. It does not show the queries or prompts that triggered the appearance. There is no API access. Data history starts from 18 May 2026, and the rollout has been incremental rather than universal.

If you cannot see the report at all, Google documents three reasons. Your property may not be in the rollout yet, the site may not have had enough AI impressions to populate it, or the site may have been excluded from generative AI features.

So it tells you that you appeared. It does not tell you whether anyone acted on it, or what they were asking when you did. Search Engine Journal has a good guide to reading the patterns in it without drawing conclusions the data cannot support. Treat it as one input rather than the answer.

A scorecard that measures the right things

If citation counts and impression totals are only part of the picture, here is what is worth measuring instead. None of it requires a subscription.

Recommendation share. When someone asks for a business like yours, how often are you the one named? This is the number that most closely tracks to enquiries.

Brand accuracy. When you are named, are the facts right? Address, services, hours, whether you are still trading. Wrong facts cancel out good visibility.

Presence. Across a fixed set of queries relevant to your sector, how often does your name appear at all, whether recommended or merely mentioned?

Build a fixed prompt set covering four types of question, and run it monthly across ChatGPT, Gemini, Perplexity and Google AI Mode. Discovery questions, the ones where someone is looking for a provider and does not know you exist. Comparison questions, where you are weighed against named competitors. Trust questions, about credentials, reviews and reliability. Logistics questions, about location, hours and how to get in touch.

Same prompts every time. What you are looking for is movement, not a single reading.

Where to fix things, in order

The research points to a sensible sequence, and it is not the order most businesses work in.

Technical eligibility comes first. Can the crawlers reach your pages at all? This sounds basic and it is routinely the actual problem. We covered it in detail in our piece on the Cloudflare setting that can block AI crawlers, and there is no point optimising content that nothing can read. It is also the sort of thing that only gets caught if somebody is looking, which is the case for a monthly maintenance package.

Trust signals come second. Consistent business information everywhere it appears, credentials stated plainly, third-party sources that corroborate what your own site says.

Content depth comes last, and it has just become more demanding. Google’s AI Mode now runs on Gemini 3.5 Flash, which carries conversational context across follow-up questions and is more selective about which pages it trusts enough to cite. Thin location pages, generic service descriptions and content that covers a topic without adding anything are at higher risk of being passed over than they were a year ago.

That last point turns a content problem into a design and structure problem. A service page that exists mainly to hold a keyword is now a visibility liability rather than a neutral asset. Depth, specifics, numbers, named people and clear statements of what you do and do not do are what earn a citation, and this is exactly the territory covered in why your website is your best or worst salesperson. If the enquiries have quietly dried up, it is usually worth starting there.

The short version

Ask whoever produces your AI visibility reporting a single question. Is this number telling me I was read, or telling me I was recommended?

If they cannot answer that clearly, the report is decorative.

If you want a straight assessment of where you actually stand across the AI platforms rather than a dashboard, our growth marketing team will run a baseline and tell you what it means, and our AI, automation and insight work picks up from there.