GEO Explained for Agencies: Why AI Visibility Is a Content Problem
Your clients are being described, compared and recommended by AI assistants in conversations they never see. This is what those assistants read before they name somebody - and why the answer is almost always published content.
What GEO means, and why the acronym barely matters
Generative Engine Optimisation (GEO) is the label the industry gave to making a business legible to AI assistants - so that when someone asks ChatGPT, Perplexity, Claude or Gemini a buying question, that business is one of the names in the answer.
The term comes from a 2024 paper by researchers at Princeton, Georgia Tech and IIT Delhi, presented at ACM SIGKDD. It is a useful piece of shorthand and a poor piece of client-facing vocabulary. You will meet it in vendor decks and RFPs, so it is worth being able to define it. We would not put it in a proposal.
What matters underneath the acronym is simpler, and it is the thing your clients actually pay for: when a buyer in your client’s category asks an assistant a question, is your client in the answer, or is a competitor?
Why this is a content problem, not a listings problem
The instinct inherited from local SEO is to reach for directories and profiles. For B2B and considered-purchase categories - software, professional services, manufacturing, specialist suppliers, anything with a long evaluation - that instinct misfires.
Assistants answering a considered-purchase question are not reading a map pin. They are reading prose: comparison pages, methodology explanations, pricing-model breakdowns, integration notes, “how do I choose between X and Y” articles, technical FAQs. They synthesise an answer out of whatever text exists on the open web that addresses the question. If nobody has written the page that answers your client’s buying question, the assistant answers from whoever did - usually a competitor, an aggregator, or a review site.
That is the whole mechanism. The gap between “my client should be recommended here” and “my client is recommended here” is, in most cases, a page that does not exist.
The Princeton study is worth knowing for exactly this reason. It tested content techniques against AI visibility and found that the things which moved the needle were editorial: adding statistics, citing credible sources, quoting named experts. Keyword stuffing did nothing. The paper’s most interesting finding for agencies is that these gains landed hardest on sites that were not already dominant in traditional search - adding citations lifted visibility substantially for a mid-ranking site, while the top-ranked site lost ground. The ordering is not simply inherited from Google. It is contestable with writing.
What assistants read before they name a vendor
Five things determine whether a client is available to be named. Only one of them is technical.
01. Whether the assistant can read the site at all
Blocked crawlers in robots.txt, a bot rule at the CDN that nobody remembers enabling, client-side-rendered content that never resolves into HTML, key claims trapped in PDFs or images. This is a fifteen-minute check that regularly explains a year of confusing results. Do it before you propose anything else.
02. Whether the buying question has been answered in prose
The substantive one. Assistants extract answers. A page that states a direct, quotable claim near a question-shaped heading is extractable; the same information dissolved through four paragraphs of positioning language is not. This is a writing standard, not a markup standard.
03. Whether the entity is unambiguous
One consistent company name, one description of what the business does, one set of facts about size, sector, footprint and offering - repeated the same way across the site, the client’s profiles and third-party mentions. Assistants build a picture of an entity from scattered sources. Contradictions make them hedge, and a hedging assistant names somebody else.
04. Whether anyone else corroborates it
A claim that appears only on the client’s own domain is weaker than the same claim appearing in trade press, an industry body’s site, a conference programme or a substantive third-party write-up. Corroboration does not require a link. A mention in the right context is doing the work.
05. Whether it is current
Assistants weight recency, some of them aggressively. A comparison page that still describes a competitor’s 2024 product line is a liability, not an asset. Refreshes are not busywork in this channel; they are a large part of the job.
Where monitoring comes in
The hard question for an agency is not how to write for AI assistants. It is what to write, for which client, this month.
Guessing is expensive. The alternative is to ask the assistants the buying questions your client’s customers actually ask, and record what comes back: whether the client is named, who is named instead, and which sources the assistant leaned on. That produces a ranked list of questions the client is absent from - which is a content brief written by the market rather than by a strategist’s intuition.
That is the first half of the loop. The second half is what happens after publication: how the articles perform, how many readers click through to the client’s own site, and when the assistants last came back and fetched the page. Those signals feed the next cycle’s decisions the same way the gaps do. You find out what the writing actually did, and write accordingly, rather than writing and hoping.
Both halves are a working process with people in it, not an algorithm running unattended. The monitoring proposes; a named editor - yours or your client’s - decides what publishes.
What this looks like as agency work
Practically, a client programme in this channel is: a monitored set of buying questions; a monthly brief drawn from the gaps and from how last month’s articles performed; articles and answer sections written against that brief; publication under the client’s brand; and a short report that re-asks the same questions and records what moved.
It is a retainer shape rather than a project shape, because the questions keep changing and the competitors keep publishing. It is also - usefully for an agency - a shape that produces a visible artefact every month.
GEO and SEO, briefly
They are not in conflict, and framing them as rivals to a client is a position you will have to walk back.
| SEO | GEO | |
|---|---|---|
| Objective | Rank in a list of results | Be named in a synthesised answer |
| Principal signals | Links, keywords, technical health | Extractable claims, citations, entity clarity, freshness |
| Content shape | Long-form, keyword-led | Direct answers to specific buying questions |
| What the buyer does | Clicks through and evaluates | Arrives already recommended, or never arrives |
| Measurement | Rankings, sessions, CTR | Presence in answers, and what the published work earns |
Most of what you would do for one helps the other. The difference is what you write, and how you decide what to write.
The short version
AI visibility in considered-purchase categories is not a directory exercise and it is not a schema exercise. It is a publishing programme, aimed at specific unanswered buying questions, kept current, and steered by evidence from two directions: what the assistants are saying now, and what the last batch of articles actually earned.
The gaps are the brief. We write what fills them - for agencies, under their brand, never in front of their client. If you place content for clients in categories like these, the partner programme is where that starts.
What to do with this
The gaps are the brief. We write what fills them.
Monitoring shows which buying questions the leading AI assistants answer with somebody else’s name. Those questions become the month’s articles, answer blocks, fact sections, FAQs and refreshes - written, published and measured under our partners’ brands.