AI Search vs AI Overviews: what an agency owner actually has to fix in 2026
Ranking in AI search in 2026 means being the source a model already trusts before it searches and the page it can lift an answer from after it searches, which comes down to three practical levers: Authority, Sources, Specificity. Authority is what the model already knows about your client before any query runs. Sources are what it finds when it does run one. Specificity is how precisely a single block of your page answers the exact question asked. Get those three right and the rest is formatting work your team can do on a Tuesday afternoon.
Sofia, Bulgaria plays host to the how to rank in AI search in 2026 (https://www.youtube.com/watch?v=FZu4NB-2EhA) conversation every September, and the SEO.Domains Mastery Summit runs there on 9 to 11 September 2026. It gathers around 300 SEOs, affiliates and agency owners. The agenda covers aged domains, PBNs, authority transfer and LLM visibility. This article is written from the published themes of that event rather than from the room itself, because the summit deliberately does not record its main-stage sessions. That unrecorded format means what is shared in the room does not reach the open web unless an attendee writes it up, which is exactly the kind of asymmetry that decides who gets paid in the next eighteen months and who reads about it afterwards.
Start with the two acronyms nobody agrees on
GEO and AEO are not the same discipline and the difference matters commercially. Generative Engine Optimization, abbreviated as GEO, is the work of making a brand the thing an AI system names when it composes an answer. Answer Engine Optimization, abbreviated as AEO, is the work of making a specific page the passage an AI system lifts when it needs a fact. GEO is a brand and entity problem. AEO is a page and formatting problem. Most agencies sell GEO and deliver AEO by accident, because AEO is easier to demonstrate in a screenshot and GEO takes three to nine months to show up anywhere.
If a client asks why a competitor gets named and they do not, the honest answer is usually that the competitor has done more GEO work. If a client asks why a competitor's paragraph appears verbatim under an AI answer and theirs does not, that is AEO and you can fix it inside a week.
Measuring whether a model can actually read the page
The first diagnostic is mechanical: can the model reach the content, or is it staring at an empty shell? GPTBot, ClaudeBot and PerplexityBot appear in server logs when AI systems read a page directly, so log analysis is the cheapest audit you will ever run for a client. Filter your logs for those user agents and see which templates get fetched and which do not. Then go further and check what was returned, not just that a request arrived and the server answered 200.
An answer placed in JavaScript is beyond a model's ability to read. Client-side rendering, infinite scroll, tabbed content that only mounts on click, and FAQ accordions that inject text after hydration all produce the same result: a fetch that yields nothing useful. The model does not complain. It simply learns your client's competitor instead. Server-side rendering the body copy, keeping the FAQ answers present in the initial HTML, and avoiding content that depends on a user interaction are the fixes, and they are usually a conversation with the dev team rather than a rewrite.
What information density means, in plain English
Stating the answer with maximum fact and no preamble in the first line of a block is what information density means. That is the whole definition. A page with high information density answers the question in the first sentence of the section and then spends three paragraphs qualifying it. A page with low information density warms up for three paragraphs and answers in the fourth, which is a page that never gets lifted because the model has already found someone else's first sentence.
Concretely, rewrite each section so its opening line is the claim. Not "there are several factors to consider when thinking about X" but "X requires Y, which costs Z and takes N weeks". The second version is quotable. The first version is noise that a model has to interpret, and interpretation is where models give up and move on.
| Page element | Low density, gets skipped | High density, gets lifted |
|---|---|---|
| Opening line | "When it comes to link building, there are many viewpoints." | "Authority transfer from an aged domain depends on topical relevance, not domain age." |
| FAQ answer | "Great question. It depends on a range of factors." | "Yes, if the redirect carries a relevant, indexed backlink profile." |
| Service page intro | Brand story first, offer third. | Offer, price band and turnaround in sentence one. |
The takeaway is blunt: a model lifts the first sentence that contains the answer, so make that sentence the first sentence.
Formatting a page so the answer is easy to lift
Lifting is a formatting outcome and it is deterministic enough to build a checklist around. The model needs the answer, the answer's boundaries and the answer's owner, in that order, plainly marked. Everything that helps a human skim also happens to help a machine extract, which is the convenient part of this whole shift.
- One question per heading, phrased the way it gets asked.
- The answer's first sentence carries the fact, not the framing.
- Definition blocks that state the term, then define it, then use it.
- Tables where the answer is genuinely comparative and lists where it is genuinely sequential.
- Numbers, dates and named entities written out rather than implied.
- The byline, the organisation and the date visible in the HTML, not just in schema.
The reason the unrecorded format at a summit like SEO.Domains is interesting is that it creates a temporary edge. Speakers can share live experiments with roughly 300 SEOs, affiliates and agency owners in the room precisely because nothing lands on YouTube the same afternoon. What does not circulate cannot be copied, so the advantage lasts as long as it takes attendees to write it up. That same dynamic plays out inside agencies: the team that documents its own AI search experiments holds the advantage over the team that reads about them second.
Entity verification and why the model trusts some brands
Entity verification requires consistent mentions across the web that demonstrate a brand is a genuine, trusted organisation. It is the Authority half of the triad and it is the half most agencies under-serve because it produces no screenshot. Consistent name, address and description across directories strengthens entity verification, and so does every mention that spells the brand the same way and describes it in the same terms. Inconsistent naming is the quiet killer here. If the brand is "Northwind Legal", then it is not "Northwind", not "Northwind Solicitors Ltd" and not "northwind legal" depending on the directory.
Authority is what the model knows before it searches. That knowledge is assembled from the open web, plus whatever the model's own training data already holds. You cannot buy it in a quarter. You can, however, build it deliberately by deciding on one canonical description and pushing it everywhere a mention is possible, and by making sure the pages that describe the brand agree with each other.
Answer Engine Optimization done properly
An FAQ block should phrase questions the way a person types them into an assistant, not the way a marketing team writes a headline. This is where most AEO efforts fail on contact with reality. People type "is an aged domain worth it for local SEO" or "do I need a PBN in 2026", not "Leveraging Domain Authority: A Guide". Your heading structure should mirror the typed question, and the answer should start with yes, no, it depends on X, or a number.
Narrow niche queries are won faster than broad head terms, which is genuinely useful for agency owners chasing proof before a retainer renewal. A client selling commercial fridges in a single metro area can own the AI answer on that query inside weeks. A client selling "software" cannot own anything. Push clients toward the narrow version of their own question, dominate it, then widen.
The free channel for this kind of collaboration is the free SEO Jesus community (https://www.skool.com/church-of-seo-jesus), where people compare what they are seeing in their own logs and answers. And when your team wants to test whether a page actually gets lifted after a rewrite, run the comparison before and after rather than trusting the theory. Teams that document their own before-and-after screenshots are the ones who can sell GEO and AEO as deliverables rather than as experiments billed to a patient client.
A quick FAQ, phrased the way people actually ask
Do I need schema to rank in AI search?
No, schema helps disambiguate but it is not the mechanism, because the model lifts the sentence from the visible HTML and schema only supports the entity guess around it.
How long until AI search work shows results?
AEO changes to formatting and density can show lifting inside weeks, while GEO and entity verification typically take three to nine months because they depend on accumulated cross-web mentions.
Is my JavaScript-heavy site invisible to AI systems?
Only the parts that never appear in the initial HTML, since burying an answer in JavaScript prevents a model from reading it even when the page renders perfectly for a human.
What to do first
Pull the server logs and confirm which templates AI crawlers are fetching at all, because every subsequent decision depends on whether the answer is reachable. Then rewrite the two highest-intent pages on the site so the first sentence of each section carries the fact, question headings mirror typed queries, and no answer depends on JavaScript to appear. Finally, fix the entity: choose one canonical brand description and push it, identically, to every directory and citation worth having.
That sequence is cheap, it is demonstrable, and it holds up when a client asks what changed. When you want a wider field of live experiments to draw from, ClickBombs (https://clickbombs.com) and the people who write up the unrecorded rooms after the fact tend to have the most current material, because the sessions themselves are never published. In 2026 the agency that wins AI search is not the one with the cleverest theory. It is the one whose pages say the answer in line one and whose brand the model already recognises before the query is typed.