AEO vs SEO: what actually changes when buyers ask instead of search

A search engine returns a list and lets the reader choose. An answer engine returns the choice. That one difference is the whole of what changes, and almost every practical consequence follows from it.

What is the actual mechanism behind an AI answer?

When someone asks ChatGPT or Perplexity for the best tool in a category, the model is not consulting a ranking. It is assembling a response from sources it retrieved, plus what it absorbed during training. The output is a synthesis: a handful of named options, usually with a sentence of justification each.

This matters because synthesis has different failure modes than ranking. A page can rank third for a term and still never appear in the answer above it, because the answer was built from four other pages that stated their claims more plainly. Ranking rewards relevance and authority. Synthesis rewards being quotable: a clear, specific, attributable statement the model can lift without hedging.

The practical test is simple. Take a page and ask whether a stranger could extract a single defensible sentence from it that names the product and says what it does better. If the answer is no, an engine will not manage it either.

Why do rankings and recommendations come apart?

Three reasons, in rough order of how often they bite.

The first is source mismatch. Engines lean heavily on pages that compare things: roundups, alternatives pages, review sites, community threads. A company's own site can rank beautifully for its brand term and appear in none of those. The engine never encounters it at the moment a recommendation is being assembled.

The second is specificity. Marketing copy is written to sound appealing; models summarise what is concrete. "Powerful, flexible and built for scale" survives no summarisation at all. "Prices per workspace rather than per channel, so adding a fifth network costs nothing" survives intact, because it is a fact with an edge.

The third is entity clarity. If a model cannot reliably tell what a product is — the category, who it is for, what it replaces — it will omit it rather than risk describing it wrongly. Products with vague positioning are disproportionately absent from answers, which is a positioning problem wearing a technical costume.

Does this make traditional SEO obsolete?

No, and the claim should be treated as a warning sign about whoever is making it.

Answer engines read the web. What they read overlaps heavily with what search engines surface. A page that cannot be crawled, that renders only after JavaScript the crawler never runs, that duplicates four other pages, or that carries no clear claim, is invisible to both systems for the same underlying reasons.

What has changed is the ceiling, not the floor. Technical SEO, clean canonicals, a real sitemap, fast rendering and sane internal linking remain the cost of entry. They no longer finish the job, because being readable is not the same as being chosen.

What does the work actually consist of?

In practice it comes down to three streams running together, and the common failure is doing one well and the others not at all.

Pages that answer buying questions. Not top-of-funnel explainers, but the comparison, alternatives and use-case pages someone reads at the end of an evaluation. These are what engines reach for when a shortlist is being assembled, because they are the pages that contain comparative claims in the first place.

Citations on the sources engines quote. If the roundups in a category never mention a product, it will not appear in answers built from them, however good its own site is. This is the part that most resembles old-fashioned PR, and the part most often skipped because it cannot be done by publishing alone.

Measurement that reflects the mechanism. Rankings and sessions do not capture whether a brand is being recommended. The honest measurement is to ask the engines the questions buyers ask, on a schedule, and record what comes back: who is named, in what order, citing which sources.

Run content without citations and the pages have no external corroboration. Run citations without content and there is nothing good to send people to. Run either without measurement and there is no way to tell which one is working.

How long does any of this take?

Technical fixes can register within weeks. New pages typically need six to twelve weeks to settle, for the ordinary reason that they have to be discovered, indexed and then actually retrieved. Citations compound slowly — a placement published in month one may not show up in answers until month three or four, because the engine has to encounter and weight it.

Anyone promising measurable movement in AI answers inside thirty days is describing something other than how retrieval works. If pipeline is needed this month, paid search is the honest answer.

Where to start if nothing is being measured yet

Begin with the baseline, because everything else is unfalsifiable without it. Write down the ten questions a buyer would genuinely ask before choosing in the category — not keywords, questions. Ask each one of the major engines. Record who gets named and which sources get cited.

That exercise usually produces two uncomfortable findings: a set of competitors who appear consistently, and a set of third-party pages doing the real work of the category. Both are the actual roadmap. The first says what the engines currently believe. The second says where that belief is coming from.

Everything after that is a matter of changing the evidence.