What is AI search, and why it matters now

01

What is AI search?

AI search is when a person asks a question and gets a written answer instead of a list of results. The tool reads a handful of pages, writes a summary, names two or three businesses and links them. The buyer reads the answer and often stops there.

That is a different competition from the one most sites were built for. Ranking well matters only if the engine reaches your page while it is assembling the answer, and only if what it finds there is quotable. A page that ranks third and says nothing extractable loses to a page that ranks eighth and answers the question in its first paragraph.

The shift is not that Google stopped mattering. It is that a second surface now sits above the results, and a third surface exists entirely outside Google in ChatGPT, Perplexity, Claude and Gemini. Your business can be strong on one and invisible on another, which is why the first useful thing to do is check rather than assume.

02

What is the difference between SEO, AEO and GEO?

SEO gets a page ranked. AEO gets its content extracted into a direct answer. GEO gets it quoted and credited by a generative assistant. They are three jobs on one foundation, and no serious version of any of them skips the foundation.

The terms are newer than the practice and nobody agrees on their edges. Wikipedia notes there is no settled definition separating GEO, AEO and the other acronyms in circulation, and at least one vendor in this market has published an argument that AEO and GEO describe the same thing. I use the split below because it is useful for deciding what work to do, not because the industry has agreed on it.

The practical consequence is that anyone selling you AEO as a replacement for SEO is selling a story. Google has published that its generative features are built on the same ranking and quality systems as ordinary Search, so the work that earns a ranking is the same work that makes you eligible to appear in an AI answer.

03

How do AI engines decide what to cite?

Nobody outside these companies knows the full answer, and anyone who tells you otherwise is guessing with confidence. What can be established is narrower and still useful: an engine can only cite a page it reached, parsed and had reason to trust, so most citation failures are actually access or clarity failures.

Access comes first. If a retrieval crawler cannot fetch the page, or fetches it and finds an empty shell that fills in with JavaScript, nothing else you do matters. This is the single most common technical cause of invisibility, and it is usually invisible to the site owner because the page looks fine in their browser.

Clarity comes second. An engine assembling an answer is looking for a passage it can lift. A page that buries the answer under three paragraphs of positioning is harder to quote than one that states it and then explains it. That is not a trick, it is the same thing that makes a page useful to a person in a hurry.

Trust comes third, and it is the slowest. It is built from being consistent about who you are everywhere you appear, being specific enough to be checkable, and being mentioned by sources the engine already relies on. A new domain is usually not in the candidate pool at all, and no amount of on-page work changes that in a week.

• Reachable: the content is in the HTML the crawler receives, not assembled afterward
• Quotable: the answer to the question is a passage, not a scatter of hints
• Checkable: claims are specific, and sourced where they are about someone else
• Consistent: the same name, description and details wherever you appear

04

Which crawlers matter, and why blocking the wrong one costs you

Each of the major AI companies runs more than one crawler, and they do different jobs. One collects content that may train a model. Another fetches pages to answer a question and cite the result. They are separately named and separately controllable in robots.txt, which means you can decline training while staying eligible for citation.

This is where a lot of sites quietly lose. A blanket block aimed at "AI scraping" usually catches the retrieval crawler too, and the retrieval crawler is the one that produces the link in the answer. The site owner then concludes that AI search does not work for their business, when what happened is that they turned it off.

The names change. They have changed several times in the last two years, and a robots.txt written from a blog post in 2024 is likely to be wrong now. Check the vendor documentation rather than a listicle, and re-check it when you next touch the file.

05

What actually works, and what does not

This is the part of the subject with the worst signal-to-noise ratio anywhere in marketing. A dozen precise-sounding statistics circulate in this market, and most of them trace back to a vendor blog citing another vendor blog. Two widely repeated figures for how often AI Overviews appear differ by a factor of three and cannot both be right.

So here is the honest version, sorted by how well each claim is actually supported. I would rather lose a sale to someone promising more than sell you a tactic I cannot defend.

The pattern in the table is worth stating plainly: the things that work are the things that would make the page better for a human reader anyway. The things that do not work are the ones that sound like a lever you can pull without improving anything.

06

How do you structure a page so an engine can extract it?

Put the answer where it can be found and make it survive being lifted. The test is simple: take any passage out of the page, show it to someone with no context, and see whether it is still true and still complete. If it needs the paragraph above it to make sense, an engine quoting it will produce something misleading, and engines tend to avoid passages that do that.

Headings phrased as questions do most of the structural work, because a heading is a strong boundary and a heading that is already a question makes the passage under it an answer. Tables carry comparison better than prose does. A short summary at the top gives both a hurried reader and an assistant the whole point before either commits to the rest.

None of this requires hiding anything or writing for machines. The page you are reading is built the way it describes, which is the only honest way to make the argument.

• A summary at the top, in real text, not behind a toggle
• One self-contained paragraph that fully answers the page title
• Headings phrased as the question a reader would type
• A table wherever the answer is a comparison
• A visible last-updated date, and sources for claims about anyone else
• Content in the HTML the crawler receives, not assembled by script afterward

07

How do you measure AI visibility?

You measure it by asking the engines the questions your buyers ask and recording what comes back. There is no equivalent of rank tracking here yet, because the answer varies between users, sessions and phrasings, so a single check is an anecdote rather than a measurement.

What that means in practice is a fixed set of prompts, run on a schedule, with the results written down: whether you were named, what you were described as, and which sources the answer drew on. The description matters as much as the mention. Being named inaccurately is a problem you can fix, but only if you know about it.

Referral traffic is the second signal and it is thinner than people expect. Assistants send far fewer clicks than search does, and much of the value shows up as people arriving already knowing who you are. Treat referral numbers as one input, not the scoreboard.

I will show you the measurement before you buy anything. If the answer is that the engines already name you and the work would not move much, that is a useful thing to learn on a free call rather than three months into an invoice.

08

What I will not claim about this

I cannot guarantee a citation. Nobody can. The engines choose, the choice varies between sessions, and no agency can prove a specific mention was caused by its own work rather than by something else that changed that month.

I also cannot promise speed. Trust signals are the slowest part of this and a new domain starts outside the candidate pool. Anyone quoting you a timeline to first citation is quoting you a feeling.

What I can do is the part that is actually within anyone control: make the site reachable and readable, answer the questions buyers really ask, keep the entity consistent, and measure where you stand so the effect of the work is visible rather than assumed. If that is not enough for what you need, I would rather say so now.