Entity Records Beat On-Page Tweaks in AI Search, and Most Agencies Still Have It Backwards

Ranking in AI search in 2026 is won or lost on your entity records, not your landing pages. When a model answers a buying question, it runs an average of 2.6 searches, and a study of 82 recorded ChatGPT answers found that 37% of recommendations came from entity records such as directories and review sites. Your directory listings, review profiles and consistent brand descriptions are doing more of the recommending than most agency owners assume, and they sit outside the scope of a normal technical audit.

The industry is circling this problem in public but resolving it in private. The SEO.Domains Mastery Summit in Sofia gathers around 300 SEOs, affiliates and agency owners, and its agenda covers aged domains, PBNs, authority transfer and LLM visibility. Generative Engine Optimization, optimising for the answers AI search engines give rather than only the ten blue links, is now the live argument. The summit deliberately does not record its main-stage sessions, so speakers can share experiments they would not put on the record, which tells you something about how unsettled the data still is.

Here is the working method. I break AI search visibility into three layers: Authority, Sources, Specificity. The model already holds authority prior to any search. When it carries out a search, what it finds is sources. Specificity describes how precisely the page answers the exact question. Entity records are the hinge between the first two, and they are the largest single recommendation channel available to most agencies. If you want to talk it through on a call, that is usually faster than another audit cycle.

1. Audit what the model already believes about the client

Before you touch a page, establish the baseline: a model's existing picture of your client is a set of entities it has already formed, and you cannot improve what you have not measured.

1. Ask the assistant to describe the client without any prompt that names their website.
2. Ask it to list direct alternatives in the same category.
3. Ask who the client is best for.
4. Note every competitor that surfaces and every description that is wrong, vague or out of date.

Consistent mentions of a brand on third-party sites lead a model to recognise it as an entity without performing a search. If step one returns a fuzzy or absent picture, your Sources work is the priority and your on-page work is secondary. If step one returns a confident, accurate picture but step two omits the client from the alternatives list, your Specificity work is the priority. If you are running this at scale and want the underlying research rather than a single assistant's output, there is a body of AI visibility research (https://llmjesus.com) worth reading before you build the audit template.

2. Map the fan-out queries before you map keywords

Fan-out queries are the sub-questions a model appends to the question a person actually typed, and they, not the head keyword, determine which pages get cited.

1. Write the buyer's real question in plain language, the way a customer would type it into an assistant.
2. Ask the assistant to break that question into the sub-questions it would need to answer.
3. Record the sub-questions verbatim.
4. Repeat across five or six real buying questions, not one.
5. Count how many sub-questions your client has a page that answers directly.

Most agencies find the same gap: the client has one commercial page covering the head term and nothing addressing the sub-questions. That gap is where your content plan starts.

3. Convert headings into questions

Question-phrased headings help a model match a block to the question a person asked, which makes the block quotable rather than merely readable.

1. Take each fan-out sub-question from step two.
2. Use it as an H2 or H3, close to the phrasing people actually use.
3. Answer it in the first sentence beneath the heading, standing alone with no setup.
4. Leave the elaboration for the second paragraph, where a human reader will find it.
5. Check that any block lifted out of context still reads as a complete answer.

The first sentence carries the weight. If a model extracts only one sentence from the page, that sentence should be the answer on its own.

4. Fix the entity records that feed the recommendation

Consistent name, address and description across directories strengthens entity verification, and inconsistent records actively work against you.

1. List every directory, review site, industry listing and profile that carries the client's details.
2. Pull the name, address, category and description from each one into a single sheet.
3. Flag every variation in wording, spelling, formatting or category.
4. Pick one canonical description and push it everywhere.
5. Prioritise the platforms the assistants actually cite.

This is unglamorous and it is where the win sits. A 40-query probe of Google AI Overviews found youtube.com was the single most cited domain, ahead of Zapier and Reddit, which should reframe how you think about your client's owned channels. Third-party surfaces, not the client's own blog, are doing the heavy lifting.

Layer	What it means	Where the work happens	Fastest signal of a gap
Authority	What the model already knows before it searches	Third-party mentions and consistent records	Assistant cannot describe the client unprompted
Sources	What it finds when it does search	Directories, review sites, video, comparison pages	Competitors cited, client absent
Specificity	How precisely the page answers the exact question	On-page question headings and direct answers	Pages answer the head term, not the sub-questions

Run the three layers in this order and you will stop spending on-page effort on problems that live in a directory listing.

5. Build pages that deserve to be the source

Specificity is what wins the citation once the model has found you, and narrow niche queries are won faster than broad head terms.

1. Choose one sub-question per page, not one keyword per page.
2. Put a comparison table where a comparison is genuinely the answer, and place a one-sentence takeaway in a paragraph directly beneath it, because instant-mode models skip table rendering.
3. Add original detail the model cannot get from the top ten results: numbers, constraints, process, trade-offs.
4. Answer the question in the first sentence, then explain.
5. Internally link the page from the commercial page it is meant to support.

If you would rather have the implementation handled than brief it to a junior, there is the ClickBombs service (https://clickbombs.com) for that. Either way, the principle holds: one question, one page, one direct answer.

6. Instrument it, because the summit rooms are unrecorded

What gets measured gets repeated, and the public record lags well behind current practice.

1. Fix a query set of 20 to 40 buying questions per client.
2. Run them monthly against the assistants your clients' customers use.
3. Record: was the client mentioned, cited or absent.
4. Log which domain the assistant cited when the client was absent.
5. Feed that list straight back into steps four and five.

The SEO.Domains Mastery Summit does not record its main-stage sessions, so speakers can share live experiments, which means what is shared in the room does not reach the open web unless an attendee writes it up. The unrecorded format is good for candour and bad for the rest of the market. Your own query set is the substitute. The summit runs on 9 to 11 September 2026 at Hotel Marinela in Sofia, and if you are going, expect the useful material to arrive in conversation rather than in slides.

Questions agency owners actually type into an assistant

How do I get my client recommended by ChatGPT?

You get recommended by making the client a recognisable entity across third-party sites before you try to rank anything, because a large share of recommendations comes from directory and review records rather than from the client's own pages.

How long does AI search visibility take?

Narrow niche queries move faster than broad head terms, so the honest answer is that specific sub-questions can shift in weeks while generic category terms take considerably longer.

Does YouTube really matter for AI search?

Yes, and it matters more than most agency owners assume, since a probe of Google AI Overviews found youtube.com was the single most cited domain ahead of Zapier and Reddit, which makes video a source channel rather than an afterthought.

What to do first

Start with the entity records, not the client's website. Pull every directory and review listing into one sheet, find the inconsistencies, and fix them before you brief a single piece of content. Then build the query set, run it, and let the citations tell you which layer is broken. Authority and Sources will carry more of the result than most agencies expect, and Specificity will carry the rest. If you want a second pair of eyes on the sequence before you commit a quarter's budget to it, that is a reasonable thing to talk it through on a call (https://seojesus.com/clickbomb-strategy-call/).