Do you actually need to do anything about AI search?

If you run a business, someone has probably told you recently that AI is coming for your website. Maybe an agency emailed you about something called answer engine optimization, or AEO. Maybe you read that you need to add a special file to your site or ChatGPT will never mention you.

Most of that advice has nothing underneath it. Some of it is flatly contradicted by the companies that actually run these systems.

So we checked. Not against other agencies’ blog posts, but against Google’s own documentation, OpenAI’s and Perplexity’s published crawler docs, a peer-reviewed paper from the ACM’s KDD conference, an academic survey that graded the whole field, and behavioural data from the Pew Research Center. Here is what we found, including the parts that are inconvenient for a company that sells this work.

Are people really finding businesses through AI now?

Yes, and it changes what they do next. When a Google AI summary appears, people click a result about half as often as when one does not.

The Pew Research Center did not survey opinions about this. It measured what people actually did, tracking the real browsing of 900 US adults through March 2025. Fifty-eight percent of them ran at least one Google search that month which produced an AI summary.

Pew also found that when an AI summary did appear, people very rarely clicked the sources it cited. The practical meaning for a business is not that ranking stopped working. It is that ranking now buys you fewer visits than it used to, because a portion of your buyers get their answer without ever leaving the results page.

One caveat we will repeat rather than bury: this is US data from March 2025, and it covers Google’s AI summaries specifically, not ChatGPT or Perplexity. We know of no equivalent measurement for other markets yet. Treat the direction as real and the exact numbers as not yet ours.

Does that mean my website stopped mattering?

The opposite. AI answers are built out of the same search index, so being findable in ordinary search is now the entry ticket to the AI layer sitting on top of it.

Google’s documentation is specific about this. For a page to appear as a supporting link in AI Overviews or AI Mode, it has to be indexed and eligible to show in Google Search with a snippet. There are no additional technical requirements beyond that.

Which means the failure mode is not exotic. If search cannot properly reach your product pages, then neither can the AI features built on top of search. The problem that keeps a business out of AI answers is usually the same boring technical problem that was already keeping it out of normal results, and it was invisible before.

Do I need to add special code or a special file to my site?

No. Google says so in writing, and this is the single most commonly sold piece of AI search advice.

On the various AI text files being marketed at business owners, Google is just as direct: you do not need to create machine readable files, AI text files, markup or Markdown to appear in Google Search or its generative features, because Search does not use them. Adding one will “neither harm nor help” your visibility.

The same guidance also rejects three more common claims:

• Chopping your content into small pieces. “There is no requirement to break your content into tiny pieces.”
• Worrying about long phrases. “You don’t need to worry that you don’t have enough long-tail keywords.”
• Buying mentions across the web. Google says seeking inauthentic mentions “isn’t as helpful as it might seem.”

Structured data is still worth having. It helps you qualify for other things in ordinary search. It is simply not the lever it is being sold as, and if someone is charging you for it as an AI visibility service, they are charging you for housekeeping.

So what actually makes an AI pick your page?

Whether your page genuinely answers the question asked, and whether it was among the handful of sources the system pulled in the first place.

This is the most reproducible finding in the research. A factorial experiment run across six different AI models and 252,000 separate trials found that relevance between the question and the document, plus where that document sat among the retrieved sources, were the primary determinants of which source got cited first.

It is an unglamorous answer. There is no format trick under it. Being picked mostly comes down to actually being the best available answer to something a real person asked.

Does the way I write change anything?

Yes, but only once your page is already in the running. That condition is the entire story, and it is the part that gets removed when this research is quoted at you.

The KDD 2024 paper that founded this field tested nine different ways of rewriting a page across 10,000 queries. Some worked well. One backfired.

Now the condition. Those gains were measured on pages that had already been retrieved and placed in a set of five candidate sources. As the 2026 survey of this research puts it, in that setup “the source is already present in a five-document context, so discoverability and much of ranking are fixed.” The study observed no clicks, no referrals, and no traffic.

So when you see “40 percent more AI visibility” in a sales email, that is this number with its condition quietly removed. The honest version is worth knowing anyway: write with real quotes, real numbers and real sources, and you improve your odds of being the one quoted, once you are already being considered. It is a tiebreaker, not a strategy.

Worth noting separately that keyword stuffing, the oldest trick in search, now actively costs you. These systems read meaning rather than counting words, and a page written for a keyword rather than a person reads exactly as badly to them as it does to your customer.

Why does all the advice contradict itself?

Partly because the engines contradict each other, and partly because the selling got ahead of the evidence.

Cross-platform audits found that Google and Gemini, two systems from the same company, overlap on well under a fifth of the sources they cite for the same questions. Different engines disagree even more, and their answers move around from day to day.

This has a direct practical consequence. Anyone who shows you a single screenshot of a single AI answer as proof of anything is showing you noise. It is the reason we measure visibility across repeated runs rather than once, and it is why we treat a one-off result as a starting point rather than a finding.

The second reason is blunter, and it comes from the survey itself: “Commercial claims about GEO have advanced faster than the evidence.”

So what should a small business actually do?

Build the page that genuinely answers what your buyer is asking, and make sure engines can reach it. It is boring, and it is the only part with evidence behind it.

We can show you what that looks like, with the limits attached. We did the ordinary version of this for an interior finishing supplier. We were the first developers to touch the site and the first to set up any measurement on it, so there was no meaningful before.

Keyword research told us which pages to build: a price list page, a local installer page, specific things specific buyers were actually typing. We built exactly those pages. In roughly five months from a standing start, that site produced more than 80 qualified inbound enquiries, covering about 6,500 square metres of quoted work. A third of the enquiries that arrived through a product page came from the long-tail pages that only existed because the research said to build them.

The honest limit on that story: it is ordinary search, not AI citation. We have not measured an AI engine citing those pages, so we are not claiming it did. We are also not claiming revenue, because we have enquiry volume and project sizes, not closed deals.

But it is the same mechanism the research keeps pointing at. Relevance to a real question is what gets a page retrieved, in ordinary search and in AI search alike. The work that makes you findable by one is very largely the work that makes you findable by the other, which is convenient, because only one of those two has a proven link to customers.

Three things worth doing

None of these need a new budget line with the word AI on it.

1. Check that engines can actually reach and read your pages. This is where the expensive, invisible problems hide, and it is a technical check rather than a marketing one.
2. Answer your buyers’ real questions, one question per page, in plain language. Write it for the person, not the algorithm.
3. If you measure AI visibility at all, measure it repeatedly. A single check tells you almost nothing.

And three things not to bother with: buying a special file for AI, paying for mentions, and stuffing keywords.