// search visibility
Answer Engine Optimization: Get Cited by ChatGPT, Perplexity and AI Overviews
A buyer asks ChatGPT which firm to use and gets three names, none of them yours. Nothing in Search Console moved that week. Citation is an inclusion decision, made per engine and per query from whatever that engine could fetch and parse at the time. That is an access problem and a sampling problem, not a position problem.
That is the boundary of this service. It covers being quoted inside generated answers and the measurement of how often that happens. It does not cover blue-link position, crawl, index or site architecture. Our technical SEO service owns those. It does not cover earning links from other sites either. That is link building. Same hub, different KPI, and you can buy one without the others.
What is actually documented, and what is not
Answer engine optimization is sold with a great deal of confidence and very little evidence. We would rather tell you which is which up front.
Not evidenced. Google states there are no additional requirements to appear in AI Overviews or AI Mode, and no special optimizations necessary. There is no schema type that makes you eligible. And llms.txt is not consumed by any production answer engine. Google’s John Mueller said so in June 2025.
Evidenced, and load-bearing. Three things:
- Access is per engine, by robots token. OAI-SearchBot for ChatGPT’s search. Claude-SearchBot and Claude-User for Claude. PerplexityBot for Perplexity. Plain Googlebot for AI Overviews, which no dedicated token controls. Google does publish a separate AI token called Google-Extended. It governs Gemini training and grounding only, and changes nothing in Search or in AI Overviews. That is exactly the confusion a wildcard rule buries. Each token is its own yes or no, and a Cloudflare bot toggle can silently reverse a decision you made in robots.txt.
- Most of these crawlers do not execute JavaScript. They fetch the script files and never run them. Content that only exists after hydration does not exist for them.
- The one peer-reviewed result. The GEO paper (generative engine optimization, KDD 2024) tested about 10,000 queries. Quotations raised visibility by roughly 41%, statistics by 33% and cited sources by 28%. Keyword stuffing lowered it by 9%. That is a single study on a synthetic engine and not a law. It is also the only real evidence in the category, and it points somewhere specific. Give figures, cite your sources, quote people by name. Putting the answer first comes from Microsoft’s published Copilot guidance, not from the study.
What the work actually is
Access
We enumerate the crawler tokens for the engines that matter to your buyers. Then the server logs and the live response tell us which ones can fetch you today. Then we fix the ones that cannot: robots.txt, CDN rules, WAF challenges, rate limits, and the platform toggles that undo all of it. None of that is glamorous, and all of it comes first: while a token is refused, every number measured after it is meaningless.
Quotability
An assistant quotes the passage that answers the question, not the page that ranks for the keyword. So we restructure the pages that should be cited. A direct answer in the opening paragraph. Claims attributed to a source a reader can check. Named quotations from people at your company who will stand behind them. Consistent facts about your entity everywhere these engines read about you. Changes ship as pull requests against your repository.
Our technical SEO service rewrites pages too, so here is the split, because no buyer should have to work it out from two sales pages. That one rewrites for crawlability, architecture and ranking: templates, internal links, canonicals, the on-site content that has to hold a position for a query. This one rewrites the same pages for extractability. The answer lifted into the opening lines, a claim a machine can quote without the paragraph around it, a source it can attribute. Buy both and you get one set of pull requests on one branch. We are not sending two people to edit the same file twice.
Measurement
A fixed panel of buyer questions, run on a schedule across engines, reported as a share of answers with the sample size attached. Next to that, the first-party reporting. Generative-AI impressions in Search Console, and the AI Performance report in Bing Webmaster Tools, which names the Copilot answers you appeared in.
Why one screenshot proves nothing
These systems are non-deterministic and personalized, and their retrieval changes without a changelog. A single run showing your brand named is one sample from a distribution nobody published. The only defensible reporting is the same questions, asked repeatedly, over time, with the count visible. If a vendor’s evidence is a screenshot, you are looking at a lottery ticket they bought with your money.
Promises we do not make
No engine sells citation share or a fixed placement, so neither is on offer here. llms.txt gets built if you want it, never billed as a lever. The GEO result rewards statistics, and we are not going to invent one to collect that reward. A fabricated number is a worse outcome than no number, for your reader and for you. Scaled pages to farm mentions are out as well. Google’s spam policies name scaled content abuse, and we treat it as a site-wide risk. And no crawler is blocked or unblocked on our recommendation until you have seen the tradeoff for that specific engine. That includes the cases where leaving an AI answer means leaving ordinary search with it.
What you get
- A per-engine access audit of which crawler tokens can fetch you today, and the robots.txt, CDN and WAF changes that fix the rest
- A rendering report, per template, separating the content that exists in your raw HTML from the content that only appears after hydration
- A fixed prompt panel of the questions your buyers actually ask, version-controlled so month-to-month readings compare the same thing
- Baseline and recurring citation readings across the engines you care about, reported with the sample count behind each reading, never as screenshots
- Rewrites of the pages that should be quotable (extractability, not site architecture), shipped as pull requests against your repository
- An entity consistency pass, so the facts about your company match across the sources these engines read
- Reporting wired to Search Console generative-AI impressions and the Bing Webmaster AI Performance report, alongside our own panel
How it runs
- 01
Establish access before writing anything
If an engine cannot fetch or parse the page, no amount of copy work reaches it. We check each crawler token and the rendering path first, because every measurement taken while an engine is blocked is invalid.
- 02
Freeze the prompt panel
The questions have to be fixed before we change the site. If the panel drifts, later readings compare different questions, and any claim of improvement becomes unfalsifiable.
- 03
Baseline with a sample size
Generated answers vary between runs and between users. One query is an anecdote, so we run the panel repeatedly and record the distribution before touching the pages.
- 04
Rewrite for quotability and ship it
The one peer-reviewed study points at statistics, cited sources and named quotations. Microsoft's published Copilot guidance adds an answer-first structure. We implement those as pull requests your engineers review, because unmerged recommendations change nothing.
- 05
Re-read, report, and say what did not move
These engines change their retrieval without notice. Re-running the same panel on a schedule is the only way to separate what we changed from what they changed, and we report both.
Questions we get asked
- How is this different from the SEO on your other pages?
- Different KPI. Technical SEO is about your own site being crawlable, indexable and understood, and it is measured in position. This is about your words appearing inside a generated answer, and it is measured in citation share. Link building is a third thing again: authority earned from other sites. The inputs overlap. The deliverables and the reporting do not. Both engagements do rewrite pages, and which rewrite belongs to which is spelled out earlier on this page.
- Is this just SEO with a new name?
- A lot of what is sold under this label is, and you should assume that until someone shows you the measurement. The honest version of the difference is this. Access is granted per engine by a separate robots token. Most of these crawlers do not execute JavaScript. The outcome is a non-deterministic inclusion decision, and not a stable ranking. Those three facts change the work enough to justify a separate engagement. Nothing else does.
- Can you guarantee ChatGPT or an AI Overview will cite us?
- No, and nobody can. Be careful with the usual version of this claim. Ads do run inside and directly beneath Google's AI Overviews and inside AI Mode. Microsoft places ads inside Copilot answers. What is not for sale is the citation itself. There is no submission form and no schema that buys inclusion in the generated answer. Google states plainly that there are no additional requirements to appear in AI Overviews or AI Mode and no special optimizations necessary. What we can do is remove the reasons you are excluded. After that we make your pages the easiest thing in the set to quote, and measure the result honestly.
- Do we need an llms.txt file?
- No production answer engine consumes it. Google's John Mueller said so publicly in June 2025 and nothing has changed that since. We will add the file if you want it for other reasons. It does not get billed as a lever or reported as work that moved anything.
- Should we block AI crawlers to protect our content?
- That is two decisions, not one, and vendors deliberately blur them. Training crawlers and search crawlers use different tokens: blocking GPTBot does not remove you from ChatGPT's search answers, but blocking OAI-SearchBot does. Google splits it a third way. Google-Extended is a real Google AI token. It governs Gemini training and grounding, and has no effect on Search or on AI Overviews. No dedicated token controls those at all. The only way out of an AI Overview is to block Googlebot, and that loses you the ten blue links in the same stroke. We will lay out the tradeoff per engine and you decide.
- Our marketing site is a client-rendered app. Does that matter here?
- More than it does for Google. Measurements of the OpenAI, Anthropic and Perplexity crawlers found they fetch JavaScript files without executing them. Content that only exists after hydration is simply absent as far as they are concerned. If that describes your site, server-rendering the content is the first fix and everything else waits behind it.
- How can you measure something that gives a different answer every time?
- We treat it as sampling, not as rank tracking. A fixed panel of questions, run repeatedly on a schedule, reported as a share of answers with the number of runs behind it. On top of that we use the first-party reporting that now exists. Generative-AI impressions in Search Console, and the AI Performance report in Bing Webmaster Tools, which shows real Copilot citations.