An answer engine optimization agency works on one outcome: being the answer a machine gives, not a link it lists somewhere below. 009 Agency does answer engine optimization the same way we do everything else — a fixed panel of your buyers’ real questions, run live through six engines, every answer and source logged, then re-run so the change is measured rather than argued.
what an answer engine is, and why the name is confusing
An answer engine is anything that responds to a question with a composed answer instead of a page of links. That definition is older than the current wave: featured snippets, People Also Ask and voice assistants were all answer engines, and the discipline of optimising for them got the name AEO years before anyone said “generative”.
What changed is how the answer is built. A snippet lifted one paragraph from one page, so the job was to own that paragraph. An assistant composes an answer from several sources it decided to trust, so the job is to be one of those sources — and to be legible enough that it can attribute the claim to you. That shift is why a second name, generative engine optimization, appeared alongside the first.
We are not going to pretend the two terms carve the world cleanly. They overlap heavily, most agencies use them loosely, and the work underneath is largely shared. Here is where they actually differ.
| classic AEO | GEO | |
|---|---|---|
| the surface | Snippets, People Also Ask, voice results. | Composed answers in ChatGPT, Claude, Gemini, Perplexity, Grok, DeepSeek. |
| how the answer forms | One passage is lifted from one ranked page. | Several sources are synthesised; some are cited, some only shape the wording. |
| what wins it | Owning the best-matched passage on a page that already ranks. | Being a verifiable entity that the sources an engine trusts already discuss. |
| what failure looks like | Someone else’s paragraph is read out instead of yours. | The model describes you accurately and still recommends three competitors. |
| how you measure it | Rank and snippet ownership for a keyword. | Share of live category answers that name you, per engine. |
what we actually do
The order below is not a preference. Entity work on a site a crawler cannot read is invisible, and citation work on a brand a model cannot verify does not hold. So it runs in sequence, and it opens and closes with measurement.
- Build the prompt panel. We write the questions your buyers actually type, not the keywords you wish they typed — category questions (“who should we pick for X”) and branded probes (“tell me about you”, “are you legit”).
- Run the baseline. Every prompt goes through six engines live. Every answer is logged verbatim with the sources it cited. This is the number everything later is compared against.
- Fix readability. Crawler access for the user-agents the engines really send, a server response that does not depend on JavaScript, and a page weight that does not time out. Half the sites we audit fail here.
- Make you verifiable. Structured data that describes you as an entity, a named team and authors, consistent identifiers across the open web, and an llms.txt that is a map rather than a blog dump.
- Make you quotable. Claims shaped so a machine can lift a clean, attributable sentence: specific, sourced, and consistent with every other page you publish.
- Get into the sources engines already trust. Which listicles, directories and communities feed answers in your category is measurable, not guesswork — the baseline told us, because it logged what was cited.
- Re-run the panel. Same prompts, same engines, and a report of the delta including the prompts that did not move.
what breaks, and how often
These are counted frequencies from our published dataset of 100 manual audits, April–July 2026 — not opinions about best practice. The full report and the raw numbers are published in the open.
| failure | found in | why the answer engine skips you |
|---|---|---|
| absent from answers in its own category | ~100% | The slot is already held by three or four names. Being good is not the qualifier; being present in the cited sources is. |
| no machine-readable schema | 94% | Nothing states what you are in a form a machine can parse, so it infers — and frequently infers another company. |
| anonymous business | 89% | No about page, no named team, no article authors. There is nothing to trust, so nothing gets recommended. |
| llms.txt missing or 404 | 87% | The one file that tells a model what matters on your site is not there. Across 100 audits we found two working ones. |
| site unreadable to AI crawlers | 50% | JavaScript-only rendering, a 403 to the crawler, or a page too heavy to finish. A human sees the product; the model gets an empty shell. |
The number that reframes the rest: across those 100 audits, models described the brand accurately in 93% of branded probes and named it in only 4.5% of “who should I pick” answers. Being known is not the problem. Being chosen is.
what this is not
It is not a prompt trick, and there is no setting inside ChatGPT that adds you to its answers. It is not a guaranteed position — anyone promising a fixed slot in a model’s output is selling something they cannot control. And it is not a replacement for your SEO: a large share of what engines cite is ranked pages, so good SEO is an input here, not a competitor.
It is also not fast in the way paid media is fast. The readability layer moves in days, the entity layer in weeks, and the citation layer as fast as other people publish. The panel is what keeps that honest, because it shows movement and the absence of it with the same rigour.