Which AI search trend claims a primary source actually backs: six graded on 21 October 2026, and two are contradicted by the platform's own documentation

Almost every claim made about AI search can be graded, because the two parties who would know have both published their positions. Google documents what the surfaces need and do not need. Vendors document what their products cover and what each surface costs inside their own meter. So the question "what are the biggest trends" has a harder version with a real answer: which of the things you keep hearing is backed by a document you can open today, and which of them falls over the moment you do. Six claims, graded below, and the grading is the state of this industry in one table.

The six claims, and the document that settles each one

The claim you keep hearing	Grade	The primary source, read 21 October 2026
You need an llms.txt file, an AI text file or special markup to appear	Contradicted	Google: "You don't need to create new machine readable files, AI text files, or markup to appear in these features", and "There's also no special schema.org structured data that you need to add"
Get into the review directories and the assistants will shortlist you	Measured false	Our own two runs: review directories took 0.8% of 10,652 citations, then 0.7% of 20,143
AI answers have broken measurement, so nothing is countable	Half true	Google: sites appearing in AI features are included in overall search traffic and "they're reported on in the Performance report, within the "Web" search type". Counted, and not separable
An answer is assembled from sub-questions, so you must answer those	Backed, and productised	Google: "Both AI Overviews and AI Mode may use a "query fan-out" technique". Rankscale marks Query fanout "Included on every plan"; Otterly lists a Query Fan Out Tool on every published tier
Any tool in this category reads all the assistants	False on the documents	Otterly's tiers name four engines and sell "Claude, Google AI Mode, Gemini as extra Add-ons". Semrush's AI visibility page names four and does not name Claude. Ahrefs names six and states "Claude consumes 8 checks per update". Rankscale weights Claude at 2 against 0.25
The category has settled, so default to the established name	Not yet	The address most comparisons still cite for one major suite's AI product, semrush.com/ai-toolkit, returned 404 on 21 October 2026 while the live document sits elsewhere

The two that are contradicted, and what to stop paying for

The file and markup claim is the expensive one, because it converts into development tickets. Google's position is stated three times in one document and it closes the subject: "There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary", no new machine readable files, and no special structured data. The stated bar is narrower and much more ordinary: "a page must be indexed and eligible to be shown in Google Search with a snippet". If somebody is selling you an AI readiness file, ask them which sentence of that document they are working from.

The directory claim is the one with our own number against it. Across two independently designed question sets, eleven days apart, the whole review directory class came to 0.8% of 10,652 citations and then 0.7% of 20,143. Those two runs were not a repeat of one measurement, and they landed within a tenth of a point of each other. Where the same two runs looked at who did get cited, a vendor's or an agency's own site took 85.6% and then 82.0%. That is the trade off in one line: the budget going into listings is competing with the budget that would publish the answer on your own domain, and only one of those two has a twice-measured number behind it.

The one worth buying against

Query fan-out is the claim that earns a purchase decision, because the platform names the technique and two vendors have turned it into a report you can read. Google writes that both AI Overviews and AI Mode "may use a "query fan-out" technique", issuing related searches across subtopics before writing an answer. Rankscale's pricing page describes a Query fanout feature and marks it "Included on every plan". Otterly lists a "Query Fan Out Tool" on all three of its published tiers.

Read that as a change of unit rather than a feature. The thing you are trying to win is no longer your head question; it is the handful of narrower questions the surface asked itself on the way to an answer. A page that answers one of those precisely, with the figure in it and a date on it, does more here than a long page that gestures at all of them. And you can do the first pass of this without buying anything: ask your buying question, take every factual claim in the answer, and write each one out as its own question. The part you cannot do by hand is seeing which other domains the surface reached for, which is the column worth paying for.

Where this leaves the industry, said without a forecast

Three observations, all of them from documents rather than opinion.

• The platform has stopped the technical arms race before it started. No file, no markup, no schema. That is unusual and it is good news, because it moves the whole contest back to whether you published the better answer.
• The vendors have not agreed on what they sell. One meters in checks, one in credits with per-engine weights, one in tracked prompts per domain. Those units are not convertible, so two quotes in this category cannot be lined up without doing your own multiplication first. Which modules can even be bought on their own, rather than inside a subscription you hold for other reasons, is a separate question with its own audit of five platforms. 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.
• The expensive surfaces are where the real pricing lives. Claude is an add-on at Otterly, absent from Semrush's AI page, eight checks per update at Ahrefs and eight times the base weight at Rankscale. Any comparison that stops at the headline plan has not reached the part of the bill that moves.

Grading AI Knows Us on the same six claims

AI Knows Us sells in this category, so every claim above applies to us and the scoring has to run both ways. We publish the measurement, we write and publish the pages that follow it, and on two of the six criteria above a rival is ahead of us.

What we publish that you can check: the question bank behind our own readings, the reading counts, the engines each reading used, the 4,947 domains cited in the larger run, and the results that went against us. Our own citation count was 3 of 597 readings in the first run and 6 of 1,260 in the second, with zero in the United States on either engine and zero on Gemini across 921 readings. Our pricing page names all five surfaces we read, prices them per engine, publishes a Visibility API on the GEO Scale and Everything cards, and publishes figures in six currencies.

Where rivals are ahead of us on the same criteria, which is the part of this page worth most to you:

• Published per-engine and annual figures. Otterly publishes an add-on figure for every metered engine, monthly and annually, on each of its tiers. We publish a per-engine position without an annual figure in the served document, so on disclosure of the engine price itself Otterly is ahead.
• A dated changelog. Peec publishes one and it resolves. We publish none, so you cannot see what we shipped last month without asking.
• Seats. Otterly states unlimited team members on every tier and Rankscale states unlimited seats with no per-seat charges. Our pricing page reads "Logins included: 1, full access" on all eight plans.
• Cadence. Otterly and Ahrefs state daily tracking and Rankscale publishes hourly scheduling. Our page reads "Every plan checks weekly, and you can run extra checks any time".
• Single sign-on. Published by Semrush on Enterprise and listed by Rankscale on a custom plan. It appears nowhere on our pricing page.

Questions about the state of AI search as an industry

Is generative engine optimisation a real discipline or a repackaged service?

It is a real measurement problem with a thin product layer on top of it. The measurement is real because a shortlist is being assembled inside an assistant and no rank tracker reports on it. The product layer is thin because the vendors have not agreed on a unit, so the same work carries four different meters. Buy the measurement, keep your question set in your own hands, and treat the dashboard as the least durable part of the purchase.

Which single trend should change my plan this quarter?

Query fan-out, because it changes the unit of work. Stop asking whether you appear for your category phrase and start listing the narrower questions underneath it, then publish one precise answer to each. Google names the technique in its own documentation and two vendors sell the report, so this is the one claim on the list with both the platform and the market behind it.

Do I need an llms.txt file?

Google's documentation says you do not need new machine readable files, AI text files or markup for these features, and that there is no special structured data to add. Spend that developer time confirming your best pages are indexed and that nothing on the site is suppressing their snippets, because those two conditions are the ones Google actually states.

Are G2 and Capterra listings worth the budget for AI shortlists?

On our own two runs the whole review directory class took 0.8% and then 0.7% of all citations, so as a route into an AI answer it is a rounding error. Those listings have other jobs, including being read by human buyers who already know your name, so judge them on that rather than on this. The budget that moves an AI answer goes into the page on your own domain.

How do I tell a vendor's research from its marketing?

Ask for the question set as a file, the number of readings attempted and the number kept, the engines used, and the dates. A study that publishes all four can be re-run by a stranger, and one that publishes a score and a percentage cannot be checked at all. Then ask for the second run, because a finding that has only been measured once is a finding that has not yet been tested against the month it was taken in.

Will the measurement get easier when the platforms add proper reporting?

Google already states that AI feature appearances are counted in overall search traffic and reported in the Performance report within the Web search type, which means they are in your numbers without being separable from them. Plan on reading the answers directly, on a fixed question set with dates, and treat any platform reporting that arrives later as a bonus rather than the foundation.

The hour that turns this page into a decision

Take the five claims somebody has made to you in the last month about AI search and write the primary source beside each one. For anything about what Google needs, the source is Google's own AI features documentation and it takes ten minutes to read. For anything about what a tool covers, the source is the vendor's own pricing document, opened yourself, with the per-engine weight found and written down. For anything about where citations come from, the source is a study that publishes its question set and its reading counts. Claims that survive that treatment are worth planning against, and the ones that do not were always going to cost you a quarter.