AI Search Optimization Services

Shane HellmrichFounder, PropeloSEO. Reviewed September 2026. Sources cited where a claim needs one.

Most AI search work fixes the wrong problem

The whole field is selling page formatting. Chunk your text, add schema, hit a readability score, publish a best in category list.

Those are on-page levers. The problem they are aimed at lives off your site.

Lily Ray put it plainly after testing it. Google appears to have decoupled what it cites from who it recommends, and the recommendation is the part that matters.

So the question is not whether a model can read your page. It is whether the rest of the web gives it a reason to name you.

The four states your business can be in

Every AI answer has two layers. The body names brands and the footnotes list sources, and they are not the same list.

That gap puts you in one of four states.

• Invisible: not named, not cited. The model does not know you exist, and nothing is being taken from you.
• Harvested: cited but not named. Your page supplies the answer and a competitor gets the credit, which is the worst state commercially and the one almost nobody spots.
• Known: named but not cited. The model trusts your brand from memory and sends the click elsewhere, which is a plumbing problem.
• Recommended: named and cited. That is the one that pays, and it is what the work builds toward.

What the work involves

Three things, in this order. Diagnosis first, then the crawlability floor, then the off-site work.

The diagnosis runs the questions your buyers ask through the models they use, then counts two things separately. How often you are named, and how often you are cited.

Not every question triggers a search. Stable factual queries get answered from training data, so prompts get classified by whether they trigger retrieval before any effort is spent on them.

The crawlability floor is blocked bots, broken rendering, dead links and slow servers. It is real and it needs fixing, but it will not make you recommended on its own.

The recommendation axis moves when other sites talk about you and link to you. Real coverage, real citations from people who are not you, and a first-page Google rank, because Ray's data is blunt that ranking is what gets you the citations.

Visibility is multi-site. A trusted review site at position three is a legitimate path into an AI answer even if your own domain never reaches page one, so the gap analysis yields content to create on your site and placements to earn on other people's.

What is not worth paying for

Some of this field is selling things that do not work. Here is what has come off the recommendation list, and why.

An llms.txt file. SE Ranking checked around 300,000 domains and found the file did not make a domain more likely to be cited, and ALLMO looked at 94,614 cited URLs across 11,867 AI answers and found exactly one llms.txt page among them.

Schema markup as a recommendation lever. It helps machines read a page and it is cheap to fix, but it is not the mechanism.

Listicles that name you the best. Writing your own best in category page has worked on language models, which is exactly why it is being cleaned up, and Ray tracked companies that scaled these into heavy declines from January 2026.

At volume, near-identical comparison pages meet Google's existing definition of scaled content abuse. Do not build them.

How this fits with the SEO you already do

The models pull from search results, so a page that cannot rank usually cannot get cited either. AI search work sits on top of ordinary SEO rather than replacing it.

It is usually cleaner to fix the crawl floor and do the off-site work than to rewrite pages that already rank. Two related reads are the GEO tools worth using and whether AI is replacing SEO.