AI search optimization

The short version

AI search optimization is the work of becoming one of the names an assistant gives when someone asks it for a recommendation. It overlaps heavily with SEO — assistants read the open web, so pages still matter — but the target is a share of answers rather than a ranking, the answer changes between identical asks, and an answer is assembled from several descriptions of you at once rather than from your homepage alone.

Everything below is the long version: what the terms mean, what actually influences the answer, how to measure it without fooling yourself, and what nobody can honestly promise you.

What is AI search optimization?

Someone opens ChatGPT and types who should I use for accounting in Manchester. They do not get ten blue links. They get two or three sentences and three or four company names.

AI search optimization is everything you do to be one of those names. In practice it breaks into four activities:

1. Measuring where you currently stand, per assistant, across the questions your buyers actually ask.
2. Finding out what the assistants read before they answer those questions.
3. Fixing how your business is described in those sources, and on your own pages.
4. Measuring again, to tell a real change from ordinary drift.

Notice that only one of the four happens on your website. That is the single biggest practical difference from the SEO most people have bought before, and it reorders budgets more than any other fact in this article.

Where did the term come from?

“Generative engine optimization” was coined in an academic paper published in November 2023, which tested how changes to source text affect whether a generative engine cites it and reported visibility gains of up to 40% from content-side changes alone.

Two things are worth taking from that paper and nothing more. First, the question is researchable: this is not a field where nothing can be known. Second, the gains measured were on retrieval-and-cite systems under controlled conditions, not a promise that rewriting your homepage moves ChatGPT. Treat the 40% as evidence that source text matters, not as a target.

The industry adopted the phrase faster than it agreed on a definition, which is why you now have four terms for roughly one thing.

GEO, AEO, AIO: what is the difference?

Do not spend time choosing a label. The practical question under all four is identical: when a buyer asks an assistant who to use, does it say your name?

We have written each out separately — GEO, AEO and Google AI Overviews — because the search demand splits along those words even though the work does not.

How is it different from SEO?

What carries over. More than the “SEO is dead” posts suggest. Assistants read the open web. Pages that were worth ranking — clear, specific, corroborated, technically reachable — are the same pages worth citing. Crawlability, structure, factual accuracy and topical depth all transfer.

What does not. Keyword-shaped thinking, rank tracking, and the assumption that your own site is the main lever. Buyers ask assistants in sentences, there is no position to track, and the deciding pages are mostly not yours. The long version of this comparison is in GEO vs SEO.

What actually decides whether an assistant names you?

Nobody outside OpenAI, Google, Anthropic and Perplexity knows how these lists are assembled, and anyone telling you otherwise is guessing. What is observable is what gets read before the answer — and that turns out to be enough to work with.

What the assistants read before answering

In one local market we measured, a single assistant answer rested on eleven pages on average. They were the businesses’ own pages alongside directories, review sites, local roundups and trade listings.

That is the most actionable number in this article, and the conclusion is not the one people expect. It is not that your own site stops mattering — it is the only description of you that you write yourself, and the one every other source copies from. Make your service pages state what you do, what it costs and who you serve, in the first two sentences under a heading phrased the way buyers ask, and you have given the other ten pages something correct to repeat.

What fails is polishing the homepage while the pages that answer a buying question say nothing specific.

Why consistency across sources matters more than polish

Assistants cross-check. A business whose services, address, opening hours or speciality differ between its own site and three directories gives a model less reason to name it than a business described identically everywhere.

This work is unglamorous. It does not photograph well in a report. It is most of the job.

Why your own pages still matter, but differently

Where your pages are read, what matters is whether they answer the buying question in the first two sentences, in the words a buyer uses. Not Our Services but Where do I go for an emergency appointment on a Sunday. Assistants extract passages; a page whose answer arrives in paragraph four does not supply one.

How do you measure AI visibility?

Why one check tells you almost nothing

Assistants do not answer identically twice. Ask the same question five times and you get overlapping but different lists. A single check is an anecdote, and building a programme on one is how agencies end up reporting weather as work.

A measurement therefore means: the same questions, asked several times per assistant, reported as a share of answers rather than a yes or no.

Why assistants are never averaged

This is the finding that surprises people most. When four assistants were asked the same set of questions about one market, they named 63 businesses between them. Only four were named by all four. Forty-six appeared in a single assistant only.

In the same measurement, shares for one business ranged from 100% of answers on one assistant down to 50% on another — identical questions, asked the same way, at the same time.

An average across four engines would have produced a middling number that described none of them, and that nobody could act on. Every figure we publish is per assistant, and the reasoning is written out in Gemini and Claude visibility.

What a usable number looks like

Named in 47% of the answers to price questions on ChatGPT. Not a rank, not a score out of a hundred, not an average across engines. Alongside it: who was named instead, and what was cited.

The questions are grouped by what the buyer wants — general recommendation, specific service, price, urgency, trust, comparison — because visibility is rarely uniform across them. A business is often named constantly in general questions and never in price questions, and that gap is the work.

The full method, including how the question set is written from the client’s own site, is published here.

What does the work actually consist of?

In the order that tends to move things, not the order that is easiest to sell.

1. Measure, before changing anything. Without a baseline, the number you get in six weeks answers no question, because compared to what? has no answer.
2. Find the question types where you are absent. Fixing a question type you already win changes nothing.
3. Identify the sources behind those answers. Where citations are shown — Perplexity always, AI Overviews usually — write them down. That list is your to-do list.
4. Get listed correctly in those sources. Accurately, completely, and identically to your own site.
5. Make your own pages answerable. Buying question as the heading, answer in the first two sentences, plain words.
6. Publish something checkable. Original numbers, stated method, real examples. Assistants prefer a source that says something specific to one that says something impressive.
7. Measure again, honestly. Movement under about five points is drift, not progress.

The expanded version of this list, with what to do at each step, is in how to improve AI visibility.

How long does it take, and what goes wrong?

Slower than it sounds. Your own pages change the day you publish them, and are read on the next crawl. The listings that repeat you move at somebody else’s pace: getting a directory entry corrected is their queue, not a sprint task.

The common failure is measuring the wrong thing. Traffic will not tell you. An assistant may name your business without linking to it, in which case the buyer arrives later by typing your name — which analytics records as direct traffic with no explanation. That is why the measurement has to be done by asking, not by reading logs.

The second failure is over-reading small movements. Answers drift several points on their own. A programme that reacts to every four-point change will spend its budget chasing noise.

What cannot be promised

Placement. The lists are not published, not stable and not for sale, and anyone offering guaranteed mentions is guessing. We say this on our own pricing page and we will say it on a call.

What can be promised is narrower and more useful: you will see where you stand, what was changed, and what moved after it — including when the answer is nothing moved, which is a real outcome and worth paying to know.

How do you know it worked?

Same questions, same way, same assistants, on a schedule. Report per assistant. Label movement under five points as no change rather than dressing it up.

If your share rose in price questions and not in general ones, that is a finding, not a rounding error: it tells you which sources responded to the work and which did not.

Who needs this, and who does not

Worth doing if your customers choose a supplier by asking around — clinics, professional firms, trades, local services, and software people evaluate before buying. Those are the categories where assistants are asked for a recommendation by name, and where being one of three names has obvious value.

Not worth doing yet if nobody asks an assistant about your category. Some purchases are made by walking past a shop, and no amount of source hygiene changes that. The cheapest way to find out which side you are on is to ask the assistants your own buying questions and see whether they name anyone at all — if they refuse to recommend in your category, there is nothing here to win.

Worth doing carefully if you are regulated. Healthcare, legal and financial categories are where assistants hedge most, and where a confident claim in your own copy can cause more trouble than the visibility is worth.

Where to start

Measure once. Everything else is guesswork until you have a baseline.

• The free AI visibility check runs a short measurement on your business and emails you what the assistants said. No call attached.
• If you would rather see the shape of a full measurement before talking to anyone, there is a complete sample report open on the site — no form, no gate.
• The method explains how the questions are written and why each is asked several times.
• Pricing is published, including what the free check covers.