Someone in your market just asked an AI which company solves their problem. It answered. It named three companies, described what each is good at, and linked two sources.

You were either in that answer or you were not. There was no page two to be on.

That is the whole of answer engine optimization, and it is why the discipline exists as something separate from search engine optimization. This guide covers what AEO is, how these systems actually decide what to surface, the strategy that works, how it relates to the neighbouring acronyms, and the shortcuts that will eventually cost you.

What AEO is

Answer engine optimization is the practice of making a company legible, credible, and retrievable to systems that answer questions directly rather than returning a list of links.

The surfaces: ChatGPT, Claude, Gemini, Perplexity, Copilot, Google AI Overviews, and increasingly the agents that book, buy, and compare on someone’s behalf.

The goal is not a ranking. It is being present in the answer, cited as the source, described accurately, and recommended for the right reasons.

How AEO differs from SEO

They share infrastructure and diverge on almost everything else.

  SEO AEO
Output A ranked list of links A synthesized answer
Positions Ten or more per page Typically one to five companies named
Stability Same result for the same query Regenerated each time, varies run to run
Unit of success A page ranking An entity being recommended
Click model The click is the goal The answer may be the entire interaction
Measurement Rank tracking, repeatable Sampling across repeated runs
Timeframe Weeks to months Weeks to months, plus training data lag

The consequential difference is the second row. Ten blue links meant a long tail of viable positions. A generative answer naming three companies means a winner-take-most distribution. Being the fourth best documented company in your category used to earn traffic. Increasingly it earns nothing.

The second consequential difference is the fourth row. SEO optimizes pages. AEO optimizes an entity, which is the model’s accumulated understanding of what your company is, who it serves, and what it is good at. Pages are how you influence that understanding. They are not the thing being ranked.

SEO is not dead and this is not that argument. Strong organic performance feeds AEO directly, because the pages models retrieve are largely the pages that already earned authority. The work compounds. It just no longer stops where it used to.

AEO, GEO, LLM SEO, AI SEO

Four names, heavily overlapping, and the boundaries are softer than anyone selling them admits.

GEO (generative engine optimization) usually emphasizes the generative surfaces specifically. LLM SEO and AI SEO are broadly the same practice with different marketing. AEO is the widest of the four, covering any surface that answers a question directly, including featured snippets and voice assistants that predate language models entirely.

We use AEO because it describes the buyer’s behaviour rather than the vendor’s technology, and buyer behaviour changes more slowly than model architecture. Pick whichever term your organization understands. Do not pay a premium for the acronym.

How AI systems decide what to surface

You cannot optimize for a black box you have not opened. Three mechanisms matter, and they have different response times.

Retrieval

Most answer engines search the live web, pull a handful of pages, and generate an answer grounded in what they retrieved. This is where the majority of near-term movement happens, because it responds to changes in weeks rather than waiting on a training cycle.

What wins retrieval: pages that answer a specific question directly and early, that are structurally parseable, that come from a domain with existing authority, and that are recent enough to look current.

What loses it: an answer buried under eight hundred words of preamble, content locked behind interaction, and pages whose actual subject is unclear from the first screen.

The practical implication is unglamorous. Put the answer first. A page that opens with the direct answer and then elaborates gets retrieved and quoted. A page that builds to a conclusion gets skipped, because the extractable passage never appears near the top.

Training data

The model’s baseline understanding of your company, formed during training and updated on the provider’s schedule. Slower to influence, harder to correct, and disproportionately powerful because it shapes what the model believes before it retrieves anything.

This is why consistency across the open web matters more in AEO than it ever did in SEO. If your positioning changed eighteen months ago but half the internet still describes the old version, the model will keep repeating the old version with total confidence, and no amount of on-site optimization fixes it.

Entity recognition

Whether the system knows you exist as a distinct thing, and what it associates you with.

Entity clarity is the most underrated work in AEO. Models need to resolve “Soulcraft” to one specific company rather than a video game, a philosophy book, or a nearby competitor with a similar name. That resolution depends on consistent naming, structured data, corroborating third-party sources, and unambiguous descriptions of what the company does.

Companies with weak entity definition suffer a specific failure: they get described in generic category language rather than in their own terms. The model knows the category and does not know you.

The four-layer strategy

Effective programs work through four layers in order. Skipping to layer three is the most common and most expensive mistake in this field.

Layer 1: Identity

Before you publish anything, you need a precise, structured answer to what the company is, who it serves, what it does better than the alternatives, and what it deliberately does not do.

This sounds like positioning work because it is. The difference is that AEO requires it in a form machines can consume: explicit, consistent, and available in a structured file rather than living in a founder’s head.

At Soulcraft this is the soul.md file. Whatever you call it, the requirement is the same. Every page, every schema block, every third-party listing traces back to one description of the company. Where descriptions conflict, models average them, and the average of an inconsistent set is generic.

Do this first. Content produced before identity is settled has to be redone.

Layer 2: Content

Build pages that answer the questions your buyers actually ask, structured for extraction.

What this means concretely:

  • One page per intent. Not per keyword. Three pages covering the same buyer question compete with each other, dilute the signal, and none of them wins. Consolidate.
  • Answer in the first hundred words. Then elaborate for the humans who keep reading.
  • Headings that are questions. They map directly to how prompts are phrased.
  • Self-contained sections. Each one comprehensible when lifted out of context, because that is exactly what happens to it.
  • Original material. Data you gathered, a method you developed, a position you can defend. Synthesis of the same twelve sources everyone else synthesized gives a model no reason to prefer you.
  • Named authors with real credentials. Attribution feeds the credibility signals these systems weight.

The last two are where most programs fail. It is possible to produce a technically perfect page that contains nothing a model could not get from four other places, and that page will not be cited, because there is no reason to cite it.

Layer 3: Technical foundation

Make the site legible to machines.

  • Schema markup. Organization, Article, FAQPage, Product where they apply. This is how you state facts in a form that requires no interpretation.
  • Clean, crawlable HTML. Content that requires JavaScript execution is retrieved inconsistently.
  • llms.txt. An emerging convention for describing your site to language models in plain markdown. Cheap to add, and it lets you frame your own content.
  • Accurate sitemaps. Live, indexable URLs only. Redirecting or noindexed URLs in a sitemap are a quality signal in the wrong direction.
  • Consistent canonical URLs. One address per piece of content, everywhere it is referenced.
  • Speed and accessibility. Same reasons as always.

Nothing here is exotic. It is the same discipline good technical SEO always required, with entity clarity weighted higher and duplication punished harder.

Layer 4: Authority

Models weight corroboration. What third parties say about you moves the needle more than what you say about yourself.

Legitimate authority work: earning coverage in outlets that cover your category, contributing genuine expertise where practitioners gather, publishing original research others cite, keeping directory and reference entries accurate, and answering real questions in public under your own name.

None of this is fast. All of it compounds. And it is where the temptation to cheat becomes strongest, which brings us to the part of this guide that matters most.

The tactics that will cost you

There is a growing industry selling manufactured authority signals. Some of it works right now. Understand what you are buying.

Self-ranking listicles. Publishing “the ten best agencies in our category” and placing yourself first. Models do retrieve these. They also increasingly discount sources where the publisher is a candidate, and the reputational exposure when a buyer notices is asymmetric.

Sock puppets and seeded forum posts. Manufactured word of mouth in communities where models retrieve. Detectable, increasingly detected, and it poisons the communities it targets.

Prompt injection. Hidden text engineered to manipulate a model’s output. This is adversarial behaviour against the platform, and platforms respond to adversarial behaviour by removing you.

Fabricated credentials and planted superlatives. Inventing awards, citing studies that do not exist, describing yourself as the leading provider in language designed to be quoted back.

Volume without review. Hundreds of pages generated and published unread. Even when nothing is technically false, the result is a site with no reason to be trusted, and it degrades the environment everyone else works in.

Our position, stated plainly: legitimate AEO makes real authority legible to machines. Structure, schema, original data, named authors, honest comparison. Faking the underlying authority is a different activity that borrows the same vocabulary, and it is a bet that detection stays behind manipulation forever. That bet has lost every previous time it was made in search.

If a comparison page cannot survive being read by the competitor it ranks below you, it is not a comparison page.

Platform differences worth knowing

The fundamentals transfer. The emphasis shifts.

ChatGPT leans heavily on both training data and live retrieval, which makes long-run entity consistency unusually important. Correcting a stale description across the open web pays off here more than anywhere else.

Perplexity is retrieval-first and cites aggressively. It rewards clearly structured, recently updated pages faster than the others, and it is the best early indicator that your content changes are working.

Google AI Overviews draw substantially on pages that already perform in organic search. Existing SEO strength transfers most directly here, which makes it the surface where traditional work pays the clearest AEO dividend.

Claude and Gemini each weight source quality and recency differently enough that you will see genuine divergence across them. This is a reason to measure per platform rather than reporting one blended number.

Do not build separate content for each. Build one excellent, well-structured, honestly-sourced page per intent, and measure where it lands. Platform-specific content is a maintenance burden that rarely earns its cost.

Measuring it

You cannot manage this without a measurement method, and the method is genuinely different from rank tracking because responses are generated rather than retrieved. Presence, citation, share of voice, and sentiment are four separate metrics with four different fixes, and blending them into one score hides the thing you needed to know.

That method has its own guide: how to measure your brand’s visibility in AI search. Read it before you start publishing, not after. It is the only way to tell whether the work is working or whether the model simply updated.

Common mistakes

Treating it as an SEO checklist. Adding FAQ schema to existing pages and calling it an AEO program. Schema helps. It does not substitute for having something worth citing.

Optimizing pages instead of the entity. Models recommend companies, not URLs. Everything on the open web that describes you is part of the surface area.

Publishing before positioning. Content produced against an unsettled identity has to be redone, and the inconsistency actively damages entity clarity in the meantime.

Keyword-shaped content. Prompts are conversational and specific. Pages built for keyword strings answer questions nobody phrases that way.

Chasing volume. More pages covering the same intent is not more coverage. It is cannibalization, and it makes every page weaker.

No original material. If a model can get the same information from four better-known sources, it will.

Measuring once. A screenshot is not a baseline. Without a frozen prompt set and a repeated cadence, you cannot separate your work from the model’s release schedule.

Getting started

If you are beginning from nothing, in order:

  1. Baseline. Forty to sixty prompts across four buyer stages, run repeatedly, on the platforms your buyers use. Record raw responses.
  2. Settle identity. One structured description of the company. Reconcile it everywhere it appears, including the places you do not control.
  3. Audit and consolidate. Find pages competing for the same intent and merge them. Most sites have more duplication than they realize, and cutting it is faster than writing anything new.
  4. Fix the technical layer. Schema, clean HTML, accurate sitemap, llms.txt.
  5. Build the pillars. One genuinely excellent page per core intent, with original material and a named author.
  6. Earn corroboration. Slowly, legitimately, in the places your category actually gathers.
  7. Re-measure. Same prompt set, same conditions, and a threshold you committed to in advance.

The programs that work are boring and cumulative. They fix the identity, cut the duplication, publish fewer and better pages, and measure honestly enough to notice when something is not working.

The programs that fail chase volume, skip measurement, and mistake activity for progress.

Soulcraft builds AEO programs for Series A through C companies: measurement, identity, and the systems that produce the content. If that is the work you need, start here.