From SEO to AEO: what actually transfers, what does not, and why

This guide is for teams that already do SEO competently and want to know what of it carries over. It is deliberately specific about what does not transfer, because that is where budget currently gets wasted.

The evidence that they are not the same thing

If AI visibility were downstream of SEO, traditional search performance would predict citation rate. In the two UAE sites we have full measurement access to, measured daily and anonymised in the report, it does not.

One site has nearly double the other's traditional search score and a materially identical AI visibility score. Two data points do not establish a general law, and we are not claiming one. But they are enough to falsify the comfortable assumption that ranking carries you into answers, and they match what we see across the wider set we track.

Why the two diverge: different units of competition

The clearest way to hold this is that the two systems compete over different objects.

That last row is the operationally important one. A ranking drop shows up as a decline you can see. Losing a buying question to an AI answer produces no signal at all — no lost click to attribute, no impression to compare against. It appears in your analytics as nothing, which is why teams routinely discover it late.

What transfers

Structured data — transfers completely

Schema markup was always partly about telling machines what a thing is rather than making it rank. That purpose is now the primary one. LocalBusiness, FAQPage, Product and Organization markup give a model facts it would otherwise have to infer from prose, and inference is where it goes wrong.

Entity consistency — transfers, and matters more

Consistent name, address, phone and description across your site, directories and profiles was a local SEO hygiene item. For answer engines it is closer to foundational. Models corroborate across sources; contradictory information does not merely fail to help, it actively reduces confidence that you are a real, established entity worth naming.

Genuine third-party coverage — transfers, and matters more

Answer engines overwhelmingly cite sources other than the subject's own website. In our source-anatomy study of one software category, the vendor's own site supplied 4.9% of the evidence engines cited. A business with no independent coverage gives a model nothing to cite about it. This is the closest thing to a durable moat in AEO, and it is also the slowest to build.

Content that answers a question completely — transfers

Long-form content built to satisfy intent rather than to hit a word count carries over well, provided it is structured so individual passages stand alone.

What does not transfer

• Keyword density and placement. Models work on meaning, not term frequency. Nothing here carries over.
• Title tags written for click-through. Curiosity-gap titles optimise a decision that never happens — there is no results page and no click to win.
• Internal link sculpting. Moving authority between your own pages does not affect whether a model considers your business a credible entity.
• Thin programmatic pages. Hundreds of near-identical location or category pages can still rank. They contribute nothing quotable and can actively dilute a thin entity footprint.
• Position tracking as a health metric. Rank one for a question whose answer box names three competitors above the link, and the ranking is not measuring what you think it is.

Search Console is the bridge

The practical problem with running both disciplines is that they produce disconnected data. Rank trackers know rankings. AI visibility tools know citations. Neither knows what the other costs you.

Search Console is the one widely available first-party dataset that joins them, because it holds the queries you already rank for and their real impression volume. Overlay citation data on it and the question changes shape:

Not "are we cited for this question", but "we rank fourth for this question, it drove 1,900 impressions last month, and the answer above us names three competitors instead."

The first is an observation. The second is a prioritised business case, and it sorts your backlog automatically — the gaps worth fixing first are the ones with real impression volume behind them. This is why we weight Search Console integration heavily when evaluating tools. It was notably rare in this category when we last compared trackers; AI visibility tools compared has the current picture.

A migration order that works

1. Measure before changing anything. Establish a baseline across every major AI engine on unbranded buyer-intent questions. Without it you cannot tell improvement from normal day-to-day variance, which is larger than most people expect.
2. Connect Search Console. Do this before optimising anything, so every gap you find arrives with impression volume attached and sorts itself by value.
3. Fix entity consistency first. The cheapest, fastest-moving work. Reconcile your business information everywhere it appears, and add or correct structured data.
4. Rewrite for extractability, not volume. Take your highest-impression pages and make each section answer one question completely in self-contained prose. Do not add pages yet.
5. Build comparison and alternatives content. Answer engines quote these heavily and most sites have never built them, because they were rarely worth it for traditional search.
6. Re-measure, then attribute. Re-run the same prompts. Movement on prompts you worked and stillness on prompts you did not is the signal that your changes caused the change.

What to expect

We are not going to give you a percentage, because we have not run a controlled study that would justify one and every number of that kind we have seen published in this category is unsourced.

What we can say from our own tracking: entity and structured data corrections show up fastest, since they change what a crawler reads on its next pass. Content changes are slower, because the source has to be re-crawled and then actually selected. Third-party coverage is slowest and most durable. Anyone promising a specific multiple on a specific timeline is guessing, and you should ask to see the method.

Questions people ask