AI Search Optimization Starts With the Questions Nobody Typed

In 2023 this field released 25 models whose category was simply "text". In 2026 it has released zero. Meanwhile 60.5% of this year's releases accept images, PDFs, audio or video, and the median model on sale holds a million tokens of context. Any AI search optimization plan that ignores those three numbers is planning for the wrong interface.

AI search optimization is usually sold as keyword work with a new name. The release data says something different: the questions themselves changed, because what people are able to ask changed. That leads to different work, and this piece sets out what.

Two ways to read the last three years

Reading one: people still want the same things, they just get them from an assistant now instead of a results page, so the job is to be in the answer.

Reading two: the questions themselves changed.

Reading one is what most of this category sells. Reading two is what the data supports.

Here is the evidence. Alongside the disappearance of text-only models, 74.1% of this year's releases name reasoning as a capability, up from none in 2023. And the median context window went from roughly two thousand tokens in 2022 to one million today.

Now hold that against the interface everyone optimised for: a text field that takes about eight words.

Every search interface ever built assumes you can put your question into words, in a box. That assumption is what keyword research measures. It has never measured the questions people could not phrase, and those questions are now askable.

You cannot photograph a problem into Google and get an answer worth having. You can into an assistant. That is not a better search engine. It is demand the search box never captured and now never will.

What this does to keyword-led strategy

Keyword tools measure typed demand. Typed demand is now a shrinking part of total demand, and it is shrinking unevenly.

I sorted six keywords into two groups before pulling any trend data. Group A: questions an assistant can finish inside the tool the user is already in. Group B: physical, local and comparison questions.

Group A fell 64.2% from peak. Group B fell 4.5%. A 14.2-fold difference. The steepest single fall was 74.7%, with no competitor involved.

So a purely keyword-led approach has a blind spot in both directions. It cannot see demand that was never typed, and it treats collapsing and stable queries as the same kind of opportunity, because the tool reports them in the same column.

What Google actually says

It is worth reading the source rather than the commentary. Google Search Central's documentation on AI features states there are no additional requirements to appear in AI Overviews or AI Mode, and no special optimisations, markup or files needed. Pages must be crawlable, indexed and eligible for snippets. Traffic from AI features appears in Search Console under the standard "Web" search type.

Two things follow from that. First, anyone selling you a special AI schema is selling you something Google says does not exist. Second, because there is no separate lever, the differentiator is content shape rather than technical configuration. The structural consequences for a working programme are covered in the five shifts AI Overviews force on a search programme.

The three jobs of AI search optimization

AI search optimization breaks down into three jobs, and most programmes only attempt the first.

1. Be a source the answer is built from. Not near it. In it.

The format evidence is specific. Of the pages that get quoted at all, Wikipedia's earn 8.11 citations each on average against YouTube's 1.87. In other words, assistants return to the same Wikipedia page for question after question, and use a YouTube page once. Dense, canonical, checkable pages get reused.

And in categories that publish prices, the vendors themselves take 90.9% of citations, against 48.7% where pricing is hidden behind a sales call.

The mechanic is consistent: what gets quoted is what can be checked. The unit of work is the passage, not the page. If a paragraph only makes sense after the two before it, it cannot be lifted, and if it cannot be lifted it will not be cited.

2. Cover the questions that were never searchable. Almost nobody does this, because no tool will hand you the list.

Get it from the places people describe problems in their own words rather than in query syntax: support tickets, sales call transcripts, community threads, the questions your solutions engineers answer twice a week. That is where "here is my situation, what should I do" lives.

3. Concentrate where a click still exists. Across a twelve-query basket, the open web now receives 0.57 clicks per search, and eight of the twelve return fewer than one. Commercial queries average 1.19; short-answer queries average 0.39. A factor of 3.1.

Winning a query that returns 0.39 clicks per search wins you very little even if you win it outright.

Four measurement rules that keep this honest

Track more than one assistant. They quote very different sources: on one site in the probe, the share of its citations coming from a single assistant ranged from 1.2% to 54.2%. And of every 100 times G2 is quoted, only about 11 come from ChatGPT. A single-assistant number is one product's house preferences reported as a market read.

Separate citation from traffic. In browsing data across 68,879 searches, clicks on sources inside an AI summary run at 1%. Citation is brand exposure – real, valuable, not a traffic channel. Why citation share is not traffic explains why the platform citing you most is often not the one sending you anyone.

Report clicks per search next to sessions. Sessions can fall while your rankings improve. Clicks per search separates the market change from your performance, which is the difference between explaining a decline and having predicted it.

Expect a visitor who is not a person. When an agent browses instead of a human, dwell time and engagement signals collapse, which breaks the measurement stack before it breaks the funnel.

Where the traffic actually is

The counterweight, because a one-sided argument is not an argument.

AI assistants now send 771 million referral visits a month, growing 117.4% year on year, converting about 54% better than non-AI sources for US retail.

It is also a low single-digit share of most sites' visits, and the reason is arithmetic: only 6.8% of assistant conversations include a web citation at all. A citation is a precondition for a referral, and most conversations never produce one.

Model AI referrals as a new channel with its own economics – roughly a fiftieth of the volume, better conversion, growing fast. Not as a replacement for the clicks that left.

Where to start with AI search optimization

If you do one thing from this piece on AI search optimization, do this.

Take your top 200 queries and sort them by a single question: can a model complete this user's task without them leaving the tool they are in?

Yes is a decline you should already be forecasting. No is where your traffic still is in 2028. It takes an afternoon and needs no new software.

Then do the harder second thing. Write down twenty questions your customers genuinely have that have never appeared in a keyword tool, because nobody could type them. That list is the real brief, and no software will produce it for you.

Everything else in this category – the tools, the dashboards, the scores – is measurement. Useful measurement, worth buying. But it measures a surface, and the surface is downstream of a change in what people are able to ask.

Questions people ask about AI search optimization

These are the questions searched most often alongside AI search optimization, answered from the data above.

What is AI search optimization? It is the work of getting your content used by AI answer engines and assistants, rather than only ranked in a list of links. In practice it combines content format, disclosure of checkable facts, and measurement across multiple answer platforms.

How do I optimise for AI Overviews? Per Google's own documentation, there is no separate mechanism: standard indexing, snippet eligibility and helpful-content rules apply. What the citation data adds is a format instruction – publish checkable facts, and structure pages so a single passage can be lifted out and still be correct.

Is AI search optimization different from SEO? It builds on SEO rather than replacing it. Pages still have to be crawlable and indexed before they can be quoted. What changes is the unit: a retrieval layer quotes passages, so passages become the thing you optimise.

What AI search optimization tools do I need? None to start. The first two steps – sorting queries by substitutability, and listing the questions nobody could type – are manual. Tools become useful when you need to measure citations across several platforms repeatedly.

How do I measure whether it is working? Track citations per platform on a fixed date and repeat monthly, report clicks per search alongside sessions, and keep AI referral traffic as its own line with its own baseline. Do not blend citation and traffic into one score.

---

Release, capability and pricing figures: a workbook of 261 model releases across 15 labs, captured 7 September 2026. Clicks per search and citation figures: live Ahrefs pulls on the same date. Click-through and source-click rates: Pew Research Center, 68,879 searches. Referral figures: Similarweb AI Search panel. Full method in the report, How AI Changed the Way People Search.