Best time to start AI search optimisation: start on the half that has been measured twice, and put the vendor decision on a short leash

It is too early to sign a long contract with a vendor in this category, and it is not too early to publish. Those two sentences look like a contradiction and they are the same finding, because the two halves of this work rest on different kinds of evidence. The publishing half rests on something measured twice and on documentation Google has kept stable. The buying half rests on vendor documents that are still being rewritten, and five of the six we opened on 21 October 2026 could not be taken at face value as they stood. Start the half that holds. Keep the other one on a short leash.

What has been measured twice, which is the part you can act on today

Our own question bank was run twice, eleven days apart in design, with two different question sets, on ChatGPT and Gemini. Run one took 597 usable readings of 612 attempted on 15 October 2026 and recorded 10,652 citations across 2,708 distinct domains. Run two took 1,260 readings on 16 October 2026 and recorded 20,143 citations across 4,947 distinct domains, which is a dataset twice the size.

What the two runs found	Run one, 597 readings	Run two, 1,260 readings
Citations recorded	10,652 across 2,708 domains	20,143 across 4,947 domains
Share taken by a vendor's or agency's own site	85.6%	82.0%
Share taken by review directories	0.8%	0.7%
Readings that cited us	3 of 597	6 of 1,260

Two question sets, designed separately, landed in the same place on the only question that changes what you do on Monday. The pages an assistant quotes in this field are overwhelmingly the vendor's and the agency's own pages. The review directory class, which absorbs a large part of many marketing budgets, came to under one per cent on both runs. The last row is ours and it is small, which is the point of printing it: the same method that produced the first two rows produced a number about us that we would rather not publish.

A finding that survives a second, larger, independently designed run is a finding you can spend against. That is what has settled in this category, and it settled on the publishing side.

What Google has kept stable, which closes off the usual reason to wait

The common reason people give for waiting is that the rules are unclear and a standard will arrive. Google's own documentation says the opposite, and has said it in the same words for long enough to plan around.

• "There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary".
• "You don't need to create new machine readable files, AI text files, or markup to appear in these features."
• The stated bar is that "a page must be indexed and eligible to be shown in Google Search with a snippet", and "There are no additional technical requirements."

So there is no file to wait for and no schema block to wait for. The entry condition is a page that is indexed and snippet eligible, which is work you already know how to do. Anybody holding off until a specification lands is waiting for something Google has said is not coming.

What is still moving, which is why the vendor decision is the one to delay

Six rival documents were opened on 21 October 2026 with two different browser user agents and the response saved before being read. Five of them could not be taken at face value as served.

Document, read 21 October 2026	What reading it found
semrush.com/ai-toolkit	404 to both agents. The live document is at semrush.com/pricing/ai/, so the address most comparisons cite for this product no longer answers
semrush.com/pricing/ai/	200, and it read cleanly from end to end. The one document on this list that did
ahrefs.com/brand-radar	200, and it marks Custom Prompts as free with a paid plan while also listing a starting monthly price for the same feature
rankscale.ai/pricing	200, and the numeric cells of its comparison table are served as 0 and filled after load, so the numbers are not in the document the server sent
peec.ai/changelog	200. The changelog is real and dated, which is more than most of this category publishes
peec.ai/trust	404 to both agents, as were /trust-center, /security and /trust-and-security
otterly.ai/pricing/	403 to the Chrome agent and 200 to the Safari agent, so whether this document exists depends on which browser asks

Note which way that table runs. A rival produced the only document that read cleanly, and a rival publishes the dated changelog. The finding is not that these are bad companies. It is that the primary sources in a category roughly two years old are still being moved, renamed and script filled, and a three year commitment signed against a document that will be rewritten by Christmas is a commitment signed against nothing.

So what has actually changed in search marketing, and what has not

Three answers, separated because they carry different weights.

1. Changed, and settled. A results page now often answers the question itself, so impressions and clicks have come apart and have to be read as two series rather than one. And a second shortlist now exists that no rank tracker reports on, assembled inside an assistant from pages it chose.
2. Changed, and not settled. Who sells the measurement, what a unit of it is, and what it costs. Vendors meter by tracked question, by recorded answer, by credit and by domain, and the per-engine weights differ by a factor of eight between the two vendors that publish one.
3. Not changed at all. The page still has to be indexed and still has to be eligible for a snippet, and the thing that gets quoted is still a specific, checkable answer on a domain somebody trusts. Two runs of our own audit put a vendor's or agency's own site at 85.6% then 82.0% of all citations, which is the oldest advice in this trade arriving with a number attached.

The order to start in, if today is the day

None of the first three steps needs a vendor, which is the practical consequence of everything above.

1. Write the questions. Five to ten questions a buyer of your product would type, with your company name in none of them. This file is the denominator for everything that follows, so date it and version it.
2. Ask them yourself and keep the answer text. Not a score, the text. The names in the answer are your real competitive set, and the claims in it are the sub-questions you have no page for.
3. Publish the specific answer on your own domain, and confirm the page is indexed and that nothing on your site is suppressing its snippet. This is the step the twice-measured finding points at, and it is the cheapest one on the list.
4. Re-read the same questions thirty days later, same wording, same market, same day of the month, and write both readings down with their denominators.
5. Only now decide whether to pay anybody, and decide it against your own question count rather than a plan name. Entry tiers in this category start low and the allowances differ widely, and the range those entry tiers actually cover is set out on its own page. That page carries no durable inbound link from anywhere on this site, which is true of 963 of our 1,987 live pages and of 175 in this category, counted across 2,075 stored bodies.

Where AI Knows Us fits in that order, said plainly

We are a candidate at step five and not before it, and the case for us is the part of the work the measurement tools leave on your desk: we write and publish the pages afterwards. Our own pricing page, read 21 October 2026, publishes per-engine pricing, names the five surfaces we read, publishes a Visibility API on the GEO Scale and Everything cards, and publishes prices in six currencies.

The honest limits on the same page matter more to somebody deciding whether to start. It reads "Every plan checks weekly, and you can run extra checks any time", where Ahrefs states a daily check per platform and Rankscale publishes hourly scheduling, so on cadence we are behind both. It reads "Logins included: 1, full access" on all eight plans, where Rankscale publishes unlimited seats with no per-seat charge. It publishes no single sign-on at all. And our own citation counts in this category are 3 of 597 and 6 of 1,260, with zero in the United States on either engine and zero on Gemini across 921 readings. A vendor that would not tell you that is a vendor whose other numbers you cannot size.

Questions about timing an entry into AI search work

Is it too early to invest in AI search optimisation?

It is the right time to start publishing and the wrong time to sign a long vendor contract. The publishing half rests on a finding measured twice across two independently designed question sets and on documentation Google has kept stable. The buying half rests on vendor documents that are still moving, so take short terms, keep the raw answer export as a condition, and re-price at renewal.

What is the cheapest honest way to find out whether this matters for my business?

Ask five buyer questions in two assistants with your name absent from the wording, and save the answer text with the date. If a competitor is named and you are not, you have your answer in an afternoon and it cost nothing. If nobody in your category is named, the shortlist is not being assembled in an assistant yet and you can revisit in a quarter with the same five questions.

When can a new page be judged for quotes in an AI answer?

It has to be indexed first, so the honest answer is as long as your indexation takes, and that is measurable on your own property rather than guessed. Submit the page, confirm it is indexed and snippet eligible, then read your questions weekly and date the first appearance. The interval between those two dates is yours to quote afterwards, and it is a far better number than any industry average.

Has the arrival of AI answers made SEO obsolete?

The opposite, on the evidence of our own two runs: the entry condition for the AI surfaces is an indexed, snippet eligible page, and the pages being quoted are overwhelmingly the companies' own. What has gone is the assumption that an impression becomes a click. Read those as two series, keep publishing the specific answer, and the same work serves both surfaces.

Should I wait for the market to settle on one tool?

Waiting for that costs you the publishing months, which are the ones with the measured finding behind them. Do the publishing now with no tool at all, keep your question file and your readings in your own hands, and buy a tool when your own question count tells you which allowance you need. A buyer who arrives with a question count and a dated baseline negotiates from a different position than one who arrives with a budget.

What should a short leash on the vendor decision actually say?

Three clauses. A term you can exit, measured in months rather than years. The raw answer text exportable at any time, so your record survives the relationship. And the question set recorded as yours, word for word and versioned, because that file is the only part of this work that cannot be rebuilt from somebody else's dashboard.

The one line that settles the timing question

If a buyer in your category can type a question today and get three names back, the shortlist is already being assembled and your absence from it is already costing you. Checking that takes an afternoon and no money. Everything else on this page is about what you do after you know, and the order matters more than the date you start: publish against the finding that has been measured twice, keep your own record, and let the vendor market finish rewriting its documents on its own time rather than on your contract.