5 Ways to Optimise for AI Search in 2026

Nov 10, 2026 · Edited by Chris Dolan

A year ago the advice for getting into ChatGPT's answers fitted on a napkin: post on Reddit, add schema, write an FAQ, get listed in a few "best of" roundups. Most of it still ranks on Google. Much of it now describes an engine that no longer exists.

In ten days of August 2026, Reddit's share of ChatGPT's citations fell by 86%. Over the same quarter, brand-owned websites went from under 2% of ChatGPT's sources to more than a fifth. Google's Gemini moved the other way. Anyone still running the 2025 playbook is optimising for last year's algorithm, and probably for only one engine.

Here are the five things the 2026 data supports, and how we approach each one at The Ranking Factory.

What changed since 2025

1. Stop optimising for "AI search" as one channel

There is no single AI search engine to optimise for. TJ Digital tracked 193,507 answers over eight weeks and found Google's AI Mode and AI Overviews shared only 12% of their cited pages for the same question. ChatGPT shared about 5% with either.

The engines also disagree about what a good source is. ChatGPT now leans on official and brand-owned pages. Google's AI still cites Reddit and YouTube heavily, and Gemini has been giving more weight to publishers, not less. A tactic that lifts you on one can be irrelevant on another.

The practical consequence is that a blended "AI visibility score" hides more than it shows. Azoma's data makes the point neatly: earned media halved on ChatGPT and rose by nearly half on Gemini in the same quarter, so a combined number would barely have moved.

How we handle it. The Ranking Factory reports visibility engine by engine and topic by topic, so you can see where you are present, where you are absent, and which engine the gap belongs to.

2. Make your own site the source of record

Your own website is now the most dependable place to earn a ChatGPT citation. In TJ Digital's data, company websites took 67% of all citations, and product pages alone took 28%, ahead of every kind of article.

The reason is a change in how ChatGPT searches. It increasingly goes straight to a named domain for a specific fact, such as a price, a spec or an integration list. HackerNoon's Q4 round-up reports that about 23% of its background searches now use a site-restricted query, up from 0.3%.

That makes the unglamorous pages matter most:

• Product, pricing and service pages that state the facts plainly. Pages showing prices had 28% higher odds of being cited.

• A clear statement of who you are, what you do and where you work. ChatGPT added a location to its searches in 30% of answers where the question named none.

• One consistent version of your facts across your site, your profiles and your listings, so the engine does not have to choose between conflicting versions.

Owned content is also the only source category you control. Engines re-weight Reddit, publishers and directories without notice. They cannot take your own pages away from you.

How we handle it. The Ranking Factory keeps a single entity record for your business, the verified facts and claims everything else is built from, and checks what the AI engines actually say about you against it. When an answer is wrong or missing, that becomes a gap to close.

3. Get onto the shortlist before the search starts

An AI engine does not search for the words the user typed. It writes its own narrower background searches first, and those decide which pages it ever sees. Studies disagree on how many it runs, from about 2 to more than 7 per question depending on the method, but they agree the searches are longer and more specific than a keyword. TJ Digital measured them at about 11 words each.

Two things follow.

Cover the sub-questions, not just the headline keyword. Pages whose opening matched the question, and whose headings matched the narrower searches behind it, were cited more often. One page targeting one phrase leaves most of those searches unanswered.

Be a name the model already associates with your category. HackerNoon reports that in eleven of thirteen product categories tested, ChatGPT added brand names to its own background searches that the user never typed. It drafts a shortlist from what it already knows, then looks for evidence. If you are not on that list, ranking first on Google does not help, because the engine never runs the search you rank for.

That association is built mostly away from your site. When an answer named a brand, a page the engine cited mentioned that brand 62% of the time in ChatGPT and 85% of the time in AI Overviews. The reliable route into the answer is to be named on the pages the engines already cite for your customers' questions.

One warning: your own "best tools" listicle is not that page. When a brand's own comparison page is cited in an AI Overview, the brand that published it is left out of the answer 70% of the time, and its competitors get the recommendation.

How we handle it. The Ranking Factory shows which sources the engines cite for your topics, and turns the ones that do not yet mention you into a prioritised list of earned actions rather than a generic outreach target.

4. Build pages that can be lifted

Pages get cited when the answer is easy to pull out of them. TJ Digital compared 21,874 cited pages with 4,507 that ranked in Google's top 10 but were passed over, controlling for domain authority and page type.

Three lessons sit behind the table.

1. Answer in the first lines. Cited pages opened with a direct answer 20% of the time, against 5% for passed-over pages. Half the passed-over pages never answered the question at all.

2. Freshness beats structure. A real update and an honest date did more than any formatting trick. Changing the date without changing the page is the kind of shortcut engines will learn to discount.

3. Schema is no longer the headline. It was the 2025 favourite and it is the smallest effect here. Do it, but after the content.

These are associations, not guarantees, and length made no difference in ChatGPT. The aim is a page a machine can quote accurately, not a longer page.

How we handle it. After a page is published, The Ranking Factory grades whether it actually answers the questions it was written for, and flags the pages that do not. That is the "does the work, not just the report" part: the platform drafts and structures the fix, and you review it before anything goes live.

5. Measure honestly, then do the work

A single check of a single prompt tells you almost nothing. Ask the same engine the same question a week later and only 24-28% of the cited pages return. Ask it twice on the same day and the figure is 23-25%. Just 2-3% of cited pages appeared in at least three of every four weekly checks.

Honest measurement in 2026 therefore means:

• Repeated runs, not screenshots. Track a fixed set of questions over weeks and read the trend.

• Unbranded questions kept separate from branded ones. Appearing when someone asks for you by name is easy. Appearing when they ask for what you sell is the number that matters, and blending the two flatters every dashboard.

• Per-engine reporting, for the reasons in point 1.

• Sources, not just mentions. Knowing which pages built the answer is what tells you whether to fix your site, earn a mention or wait out an engine change.

Measurement is only half of it. The other change since 2025 is that volume has stopped working. Publishing hundreds of thin pages or seeding forums was always a poor idea. Now that the engines have shown they will drop an entire source category in ten days, it is a liability too. What holds up is a smaller amount of accurate, current, well-structured evidence about what your business actually does.

How we handle it. This is the gap The Ranking Factory was built for. Most AI visibility tools stop at the report. Ours finds the gaps, prepares the content and evidence to close them, holds everything for your approval, publishes it, and then checks whether visibility moved. It also keeps doing ordinary SEO, because ranking in Google's top 3 still lifts your odds of an AI Overview citation from 8% to 13%.

You can see where your own site stands with the free site check on our homepage.

Sources

• TJ Digital, 2026 AI Citation Study (25 October 2026): 193,507 ChatGPT, AI Mode and AI Overview answers for 78 businesses, 27 August to 21 October 2026.

• Azoma, Why AI visibility drops (5 October 2026): more than 38 million citations on commerce prompts, 1 July to 14 September 2026.

• Stacy Muur, AI Search Optimization in Q4 2026, HackerNoon (29 October 2026): a round-up of third-party research, including the Promptwatch Reddit figures and Lily Ray's analysis of ChatGPT's background searches.