GEO: be the brand AI engines recommend

How AI answer engines actually retrieve and cite

Most AI answers are not recalled from training. They are retrieved at question time. The system takes your question, rewrites it into several related search queries, runs them against a search index, reads a set of returned pages, and writes an answer from the passages it found. The citations you see are the sources it read, not a ranked list.

Google calls its version query fan-out and describes it in its own announcement and developer documentation, so the mechanism is confirmed rather than inferred. The practical consequence is that one visible question becomes many hidden ones. A buyer asking who is the best provider for a task may trigger sub-queries about pricing, process, alternatives, credentials and common problems. Our reading of that is that covering the adjacent sub-questions gives you more chances to be one of the pages read. That last part is our inference, not an instruction from Google.

This also explains why citation is unstable. Two people asking the same question minutes apart can get different sources, because the fan-out queries differ and the retrieval set differs. Anyone promising fixed AI rankings is describing something that does not exist.

• The question is expanded into several sub-queries before anything is retrieved
• Pages are read at answer time, so being findable now matters more than being memorised
• Citations reflect what was retrieved and used, not a stable ranking
• Covering adjacent sub-questions is our reasoned bet, not a documented rule

Why entity consistency decides whether you get named

An answer engine has to be sure who you are before it will recommend you. It builds that picture from every mention of your name it can find: your site, your profiles, directories, review sites, press coverage, and any database that holds business records. If those sources disagree, confidence drops and a competitor with a cleaner record gets named instead.

Disagreement is usually mundane. Three versions of the company name across four profiles. An old address on a directory nobody updated. A services list that says one thing on the site and another on LinkedIn. Two different founding dates. None of it looks important on its own, and together it makes you harder to describe with confidence.

Entity work is the least glamorous part of GEO and often the highest value. We settle on one canonical description of the business, apply it everywhere we control, correct it everywhere we can reach, and make the site itself unambiguous about what the organisation is, where it operates, what it sells and who is behind it.

• One exact legal and trading name, used identically everywhere
• Matching address, phone and service list across every profile
• Named people with real credentials and stable bios
• Clear organisation markup on the site for entity clarity, not for citation
• Old profiles you no longer use either updated or removed

What makes a passage quotable

A quotable passage is one that survives being cut out of your page and dropped into an answer with no surrounding context. That is a writing problem more than a technical one. It means the answer comes first, the sentence names its own subject rather than relying on a pronoun, and any qualifier that changes the meaning sits inside the same passage.

The evidence on what helps is directional and worth stating carefully. A peer-reviewed benchmark from 2024 found that citing sources, including relevant statistics, and quoting credible authorities improved how much of a source appeared in generated answers. The same benchmark found keyword stuffing performed worse than the plain baseline. That study measured word share on a test set, not live traffic, so we treat it as a strong hint rather than a rule.

What we take from it in practice: be specific, attribute your numbers, write plainly, and drop the promotional voice. Marketing copy is hard to quote because it makes claims a system cannot verify. A page that says what something costs, how long it takes and what usually goes wrong is far easier to lift than one that says you deliver excellence.

• Direct answer first, detail second, in every section
• Self-contained sentences that name their subject
• Specific numbers, dates and named sources instead of vague superlatives
• Plain, non-promotional tone throughout
• Content present in the raw HTML, since several major AI crawlers do not run JavaScript

The honest limits of what anyone can influence

No agency can guarantee a citation. Citation is a long tail: the top ten domains account for only around twelve percent of ChatGPT citations, which means the field is wide and mostly made up of sources nobody would have predicted. No one can attribute a specific citation to a specific piece of work, because there is no reporting layer that shows why a source was chosen.

Some popular tactics do not survive checking either. llms.txt is the clearest example. Google has stated no AI system currently uses it. A log study across roughly 137,000 domains found that 97 percent of published files received zero requests in a month, and AI retrieval crawlers made up around one percent of the requests that did happen. A separate study of about 300,000 domains found no correlation with citation at all. We will publish the file if you want it, but we will not tell you it does anything.

Schema is the second example. The measured study of pages adding JSON-LD found no meaningful lift in AI citation. Schema stays in our work because it earns rich results and clarifies your entity, which are real benefits with real evidence. It is not a citation lever.

Why third-party coverage outperforms your own pages

A 2025 study found AI search systematically favours authoritative third-party sources over brand-owned and social content. That fits how these systems are designed. When a system is asked which provider is best, a page written by that provider is the weakest possible evidence, and a review site, trade publication or independent comparison is much stronger.

This is the single most important strategic point in GEO, and it is uncomfortable, because it means the work extends past your own website. Being present and accurately described in the places that get retrieved matters more than another blog post on your domain. Industry directories, comparison sites, review platforms, trade press, podcasts and genuine expert commentary all feed the same pool.

It also means on-site work alone cannot honestly be sold as a route to citation. We do the on-site work because it is necessary and because it makes you eligible. We pair it with digital PR and earned coverage because that is what the evidence says moves the needle.

• Independent reviews and comparison pages carry weight your own pages cannot
• Being listed accurately on the directories your sector actually uses
• Named expert commentary in trade press and podcasts
• Correcting inaccurate third-party descriptions of your business

How to measure AI visibility without inventing numbers

There is no Search Console for AI answers, so measurement has to be built. We define a set of questions a real buyer would ask, run them across the major assistants on a fixed schedule, and record what comes back: whether you are named, whether you are cited with a link, which competitors appear, and whether the description of you is accurate.

Repeat the same prompts in the same way each time. Answers vary between runs, so a single check tells you almost nothing and a monthly series tells you the trend. Share of voice against a named competitor set is the most useful headline, because it is relative and therefore less affected by the noise in any individual response.

We pair that with what analytics can see. Referral traffic from AI tools is now identifiable for several of them, and it usually arrives in small volumes with unusually high engagement. Branded search volume is the other signal worth watching, because AI answers often create a name to search for rather than a click.

• A fixed prompt set, run on a schedule across the main assistants
• Presence, citation and accuracy logged separately for each run
• Share of voice against a named competitor list
• Referral sessions from AI sources in analytics
• Branded search trend, since many AI answers create a search rather than a click

Correcting what AI tools say about you

A common reason businesses come to us is not absence but inaccuracy. The assistant names them and gets it wrong: an outdated service list, a location they left years ago, a merger described incorrectly, or a confusion with a similarly named company.

There is no correction form. The only reliable route is to change the source material the systems read. That means fixing your own site first, then working outwards through every profile, directory and article that carries the wrong information, in rough order of how likely each is to be retrieved. Where a third party will not update, the counter is to publish a clearer, better-sourced version elsewhere so the accurate account is the one more easily found.

It is slow work and it does not resolve on a fixed timeline, because it depends on when those sources are re-crawled and re-read. What we can commit to is the method, the sequence and the monthly record of whether the description is improving.