Answer engine optimization best practices, sorted by the evidence behind them

Answer engine optimization best practices, in short

Let the assistants read you, answer the buyer’s question plainly with evidence behind it, and get named by other sites. Almost everything else on most best-practice lists is either a restatement of those three or untested.

Answer engine optimization, or AEO, is the work of getting your brand named and your pages cited in answers from AI assistants such as ChatGPT, Perplexity, Gemini and Google’s AI Overviews. Our page on what answer engine optimization is defines it, and our plain guide to AI search optimization explains how the answers are built. This post is narrower. It takes the practices people recommend and asks how we know each one works.

How each practice is graded

Each practice gets one of three grades, based on the kind of source behind it. The grade says how sure you can be, not how much it will help.

Most lists of AEO tips skip this step and present everything with the same confidence. That is how myths spread. A tactic that one vendor guessed at gets repeated until it reads like a rule.

Access practices: let the assistants read you

These have the strongest evidence of any practice in this post, because the providers document them. If you do only one thing, do these.

1. Allow each assistant’s search robot in robots.txt: OAI-SearchBot for ChatGPT, PerplexityBot for Perplexity, Claude-SearchBot for Claude, and Googlebot for Google’s AI answers.
2. Check your firewall or bot filter as well. A block there undoes an allow in robots.txt, and Perplexity’s crawler page gives firewall steps for exactly this reason.
3. Set the training robots separately and on purpose. GPTBot, ClaudeBot and Google-Extended are about training, and the providers keep them apart from search.
4. Make sure key pages are indexed by Google and are not limited by a stray nosnippet or a small max-snippet setting.
5. Put the answer in the page’s text, not only in an image, a PDF or a script that runs after the page loads.

The grades for these are mostly Documented. OpenAI’s crawler page says sites opted out of OAI-SearchBot “will not be shown in ChatGPT search answers”. Perplexity’s crawler page recommends “allowing PerplexityBot in your site’s robots.txt file”. Anthropic’s crawler page says blocking Claude-SearchBot “may reduce your site’s visibility”. Google’s generative AI guide says a page must be “indexed and eligible to be shown in Google Search with a snippet”. The last item, text over images, is Reasoned: a robot that reads text cannot quote a picture of it.

Our engines page lists the robot names for ChatGPT, Perplexity, Gemini and DeepSeek, so you can check yours against one list. Our page on AI crawler access explains how a rule written for every robot, or one a plugin added, can block an assistant without anyone meaning to.

Page practices: be the most checkable answer

One study measured that sources, quotations and statistics raise how much of an AI answer comes from a page. Google’s guidance adds that first-hand, original content matters most over time.

The GEO paper by Aggarwal and others, published at KDD 2024, rewrote pages in nine ways and measured the effect in AI answers. Adding citations, quotations and statistics gave a relative gain of 30 to 40% on its main measure. The authors also found that lower-ranked sites gained more than top-ranked ones. Their test engine read five pages per question with an older GPT-3.5 model, so treat the size as a direction.

Google’s guide says unique, useful content “will likely influence your website’s presence in generative AI search in the long run more than any of the other suggestions”. It warns against commodity content that “simply restates information already available elsewhere”. Google’s page on helpful, people-first content asks whether a page provides “original information, reporting, research, or analysis”, and whether pages “carry a byline, where one might be expected”.

Answering in the first lines is graded Reasoned on purpose. It is the most repeated AEO tip, and it makes sense: a model writing a short answer can lift a short, clear sentence. But nobody has published a test of it alone, and Google says there is “no ideal page length”. Do it because it helps readers too, not because it is proven.

What a checkable claim looks like

The product and the details in this example are made up to show the difference. They do not describe any real company.

A page about an invoicing tool says: “Fast setup and great support for growing teams.” Nothing in that sentence can be checked, and nothing in it answers a buyer’s question. A model has no reason to quote it, because it would say the same about any product.

The rewrite says: “Setup takes about 20 minutes for a team of up to 10 people, based on the median of our last 200 sign-ups in 2026. Support replies by email within one working day.” Each claim now has a number, a scope and a source a reader could ask about. That is the kind of change the GEO study tested, though its effect on any one page cannot be promised.

Give every claim on the page something a reader could check: a source, a number, a date or a name. That is the practice with the clearest measured effect.

Off-site practices: be named by others

A 2025 preprint found AI search leans strongly towards third-party sources over a brand’s own site. So being named fairly on reviews, comparisons and lists is part of the job, not an extra.

The Chen preprint, which has not been peer reviewed, compared AI search with Google across many topics, languages and rewordings. It reports “a systematic and overwhelming bias towards Earned media (third-party, authoritative sources) over Brand-owned and Social content”. It also found that AI search services differ from each other in freshness and sensitivity to phrasing, which is a reason to check each assistant on its own.

• Find the third-party pages the assistants cite for your buyers’ questions, by asking the questions and opening the links.
• Where you genuinely belong on one of those pages, give its author the facts they need to judge you fairly.
• Keep your listings on review sites and directories accurate, so what they say matches your own pages.
• Do not buy or seed fake mentions. Google says inauthentic mentions are not as helpful as they might seem, because its spam systems sit underneath its AI answers.

The first three items are Reasoned, built on the Measured finding above. The last is Documented by Google for its own search.

Measurement practices: know whether it worked

Ask the same buyer questions on a schedule, keep every answer whole, and count separately the questions that name you and the ones that do not. Without this, every other practice is a guess.

1. Write 20 to 50 questions a buyer asks before choosing, in the buyer’s words.
2. Mark which questions name your brand. Those usually get your name back and say nothing about new buyers.
3. Ask every question in each assistant you care about, and save the full answer, its links and the date.
4. Say what every figure is out of. Write “named in 6 of 120 answers”, not a bare percentage.
5. Ask again on a schedule, and compare question by question, because a single answer can change on the next run.

The numbers in item four are a made-up example of the format, not a measurement. Our guide to reporting AI visibility honestly explains why the denominator, the number a figure is divided by, must always be on the screen.

Agentcites runs this loop. Today it asks ChatGPT, Perplexity, Gemini and DeepSeek up to 50 buyer questions for each website every week, stores every answer word for word with its links, and keeps branded and non-branded questions apart. It costs $149 a month per website, as the pricing page shows.

Practices to drop

Some popular tactics have evidence against them, or a provider says it ignores them. Stop spending time on these first.

Two of these need a fair word. The llms.txt proposal is aimed mostly at agents reading documentation, and Google’s guide says keeping one is “completely fine” for other services that use it. And structured data is not useless: Google’s structured data policies say only that you must not “mark up content that is not visible to readers of the page”. Used honestly, it still helps with rich results in ordinary search.

What a best-practice list cannot tell you

No list can say which practice will move your own answers. The providers do not publish their ranking, and the studies tested their own questions, not yours.

So treat this list as an order of work, not a promise. Fix access, which the providers document. Improve the page in the ways research measured. Earn mentions elsewhere. Then measure your own questions, change one thing at a time, and believe your own before-and-after comparison over anyone’s list, including this one.