The AI search visibility guide: get cited by AI

Usually not you. Here is a real case. "best GEO tools" is a question Attensira tracks for its own brand, and one we lose on both engines. From 4 September to 3 October 2026, ChatGPT answered it 57 times with 1,884 citations, and Perplexity answered it 57 times with 579. Attensira ranks the pages behind each answer by how much they decided it, citation share weighted by position, not by raw count, read here over the Attensira MCP on 4 October.

The top four per engine are the frame at the top of this page. On ChatGPT they are xseek.io, a roundup of answer engine software, then bestgeosoftware.com, sitepoint.com and geosoftwarerankings.com, all three roundups. On Perplexity they are geoptie.com, a roundup of GEO tools, bestgeotool.com, a listing site, visby.ai, a roundup, and producthunt.com, a category page. Ranks five and six are in the table below.

• 5

  ChatGPT's deciding sources

  elmohq.com, a competitor's roundup

  Perplexity's deciding sources

  seranking.com, a competitor's roundup

• 6

  ChatGPT's deciding sources

  frictionai.co, a roundup

  Perplexity's deciding sources

  nogood.io, a roundup of enterprise tools

Our site is on neither list, and we have no verified placement on any of the twelve pages. Three lessons follow, and they are the shape of this whole guide.

The deciding pages are lists other people wrote. All twelve are roundups, listing sites or category pages. Not one is a vendor's product page. For a "best X" question, the engine is mostly summarising other people's comparisons.

The two engines lean on different lists. The overlap between the two columns is zero. Winning this answer on ChatGPT and on Perplexity are two separate jobs.

Raw counts mislead. On ChatGPT, bestgeosoftware.com carried 14.3% of all the citations, but xseek.io ranks first because its link sat higher in the answers. A plan built on the most-cited page would chase the wrong one.

The same pattern holds across everything we track: the answers to our 61 questions carried 97,149 citations in 30 days, and 1,442 of them pointed at attensira.com. Being visible in AI search is as much about other people's pages as about your own.

The share of AI answers to your buyers' questions that name your brand or link your site, read per engine and reported with the number of answers behind it. It is a measurement, not a score, and it only means something relative to the questions it was measured on. The AI search visibility page goes through which numbers can be measured honestly and which are constructed.

This guide is about changing that measurement. The order matters. A page an engine cannot fetch cannot be cited, however well it is written. A well-written page that no trusted source corroborates may be read and passed over. And none of it can be judged without reading the answers before and after. So the guide follows the path an answer takes, from fetch to citation, and ends with measurement and with what to stop doing.

Four conditions, in order. Each one is a chapter of this guide.

1. The engine can fetch your pages. Its search agent is allowed by your robots.txt, the page loads quickly, and the answer is in the HTML rather than assembled by a script. How AI assistants find your pages covers each engine's agents.
2. Its retrieval finds them for the question. The engine searches its index, or a classic search engine, for pages that might answer. A page that ranks somewhere helps; a page that answers the exact question and its sub-questions helps more. Ranking and being cited explains why the two diverge.
3. A passage answers the question on its own. Engines build answers from passages. A passage that depends on the paragraph before it, or never says your name, is hard to use. Passages that survive extraction shows the difference.
4. Other sources the engine trusts agree. The table above is this condition. Why most of your citations are not on your domain covers how to earn a place on those pages honestly.

Two more chapters cover the technical work that matters, and the schema that does not, and measuring whether any of it worked. The last chapter is what does not work.

Through their own search agents and through classic search engines. Each operator documents separate agents for separate jobs. OpenAI's OAI-SearchBot surfaces websites in ChatGPT's search features, and OpenAI says a site that disables it will not be shown in ChatGPT search answers; GPTBot, by contrast, collects training data. Anthropic runs Claude-SearchBot for search quality and Claude-User for fetches when a person asks Claude something. Perplexity's PerplexityBot surfaces and links sites in its results and is not used for training. Google's AI Overviews and AI Mode sit on Google Search itself: a page must be indexed and snippet-eligible to be a supporting link.

Two consequences matter most. Blocking the wrong agent removes you from that engine's answers while you think you only opted out of training. And because each engine keeps its own index on its own schedule, a fix reaches the engines at different speeds. Google says crawling can take from several days to several months depending on how often it decides a page needs refreshing. The AI crawler guide is the long version of this section.

Because the engine is not reading the ranking. It reads pages it retrieved, possibly with several searches behind one question. Google says AI Overviews and AI Mode may use a query fan-out technique, issuing multiple related searches across subtopics and sources before writing a response. A page that ranks first for the head term can lose to pages that answer the sub-questions the engine actually asked.

Then the engine chooses passages, not pages. A ranking page whose answer is buried halfway down, or phrased in category jargon, can be retrieved and passed over. And for comparison questions, the engine may prefer a page that compares several products to any single vendor's page, which is exactly what our table shows.

A passage an engine can quote alone. The research on generative engines tested nine ways of editing a source page and measured which made it more visible inside generated answers. Adding quotations, adding statistics and citing sources were the strongest. Keyword stuffing did worse than leaving the page alone.

In practice that means:

• Answer first. The opening sentences under a question-shaped heading answer it completely.
• Self-contained sentences. No "as mentioned above", no pronouns pointing at a previous paragraph.
• Your name inside the answer. If the sentence that answers the question does not say who it is about, the engine can use the fact without crediting you.
• Specifics with sources and dates. Numbers, named sources, and when a fact was last checked.

The answer passage extractor shows which passage of a page an engine is likely to lift, and answer engine optimization goes deeper on writing them.

Honestly, one page at a time, starting with the ones engines already trust. Attensira's off-site work runs in a fixed order: first the sources an engine already cites for your questions that name a competitor and not you, then sources that rank above you but have not been read citing you, and only when nothing worth joining exists, a new original page worth citing.

For the table at the top of this page, that means contacting the authors of the roundups with accurate information, listing the product where listings are open, such as category directories, and publishing a comparison of our own that is fair enough to be useful to the people writing roundups. It does not mean paying for placement or writing fake reviews. Engines and readers both discount pages that read like advertising, and a placement that misstates what a product does is a liability the next time someone checks.

Take the question from the top of this page and walk it through. It is the method Attensira's crew follows, and you can follow it by hand.

Start with the engine's own list, not your guess. The table is ranked by how much each page decided the answer. On ChatGPT, the first page carries a small share of citations but sits high in the answers; on Perplexity, a listing site and a category page sit in the top four. Neither list is what a brand would guess from its own SEO tools.

Split the list by what you can do on each page. For a roundup, the move is a factual note to the author with what the product does and what has changed, and a request to be considered at the next update. For a listing site or category page with open submissions, the move is an accurate listing. For a competitor's own roundup, there is usually no move; it tells you which framing of the category the engine is absorbing.

Pick one next action per engine, not a list. Attensira's win plan does exactly this. For "best GEO tools", with no evidence yet from our own workspace, it falls back to its standard table and suggests our own page on ChatGPT and being in the threads and roundups on Perplexity. One action per engine is a plan that gets done; a list of twelve is a plan that does not.

Date the action and plan the re-read. Each action has a date, and the answers are read again after it lands. Only then do you know whether that page decided anything for you, which is the evidence that tells you whether to do the same on the next question.

The ones closest to a purchase that you currently lose, where the deciding pages are reachable. A short way to choose:

1. List the questions where a competitor is named and you are not. These are lost answers with a known winner, which means the deciding pages are probably comparisons you can join.
2. Drop the ones where every deciding page is a competitor's own site. There is little to work on there in the short run.
3. Prefer questions where your own page is already linked but not named. The passage rewrite is the cheapest win in this field; the engine already reads you.
4. Prefer questions asked close to a decision, comparison and "best for" questions, over general education questions, unless education is how your buyers find you.
5. Keep the list short. Five questions worked properly beat thirty touched once.

The conditions are the same; where the deciding pages live is not.

Software companies see comparison roundups, review sites, documentation hubs and community threads decide most buying questions. Their own docs and pricing pages are often read, which makes passage work unusually effective. A pull request against the company's own docs is often the fastest fix.

Professional services see directories, review platforms and local listings decide "who should I hire" answers. The passage that matters may be a listing description rather than the website.

Ecommerce brands see product roundups, retailer pages and video reviews decide "best X for Y" answers. Product descriptions that put the use case and the distinguishing fact in one sentence are the passage work.

Agencies running several brands face all of the above, per client. The method does not change; the volume does, which is where software that reads answers daily and drafts the work for each brand earns its keep.

Yes, and it is mostly the same technical SEO. Google says there is no special schema.org structured data to add for its AI features and no AI text file to create. What matters is the old floor: pages that can be crawled, indexed and shown with a snippet, fast responses, server-rendered text, and no accidental blocks in robots.txt or snippet controls such as nosnippet.

Structured data is still worth having where it describes the page accurately, because it helps engines understand what a page is about. It is not a shortcut. A page with perfect markup and no passage that answers the question is still not cited.

By reading answers, per engine, on a fixed set of questions, before and after each change. Count how often you are named and how often your site is linked, keep the sample size beside every number, and treat a change as real only when it clears a significance test. Search Console helps for Google only: Google reports traffic from its AI features inside the Web search type of the Performance report, mixed with classic results. It says nothing about ChatGPT, Perplexity or Claude.

Expect clicks to be a weak signal. Pew Research Center found that people clicked a link inside a Google AI summary in 1% of visits to pages showing one. Being named in the answer is the outcome; a visit is a bonus. The AI search visibility page covers which numbers can be measured honestly, and the GEO measurement guide covers how to measure them.

A sequence that does the cheap, certain work first.

Weeks 1 and 2: access and baseline. Check robots.txt against every engine's search and user-triggered agents. Confirm key pages render their text on the server. Write the question set, twenty to fifty buyer questions without your brand name, and start reading answers on every engine you care about.

Weeks 3 to 6: the passages you already own. For questions where an engine links your page but does not name you, rewrite the passage. For questions where no page of yours answers, write the page. Keep each change dated.

Weeks 5 to 10: the pages that decide. For the questions you lose, list the deciding sources per engine. Work the ones that already cite competitors: corrections, listings, pitches, contributions.

Weeks 8 to 13: read and adjust. Compare each engine's answers with the baseline, question by question, with sample sizes. Keep what moved, drop what did not, and say "nothing we can prove yet" where that is the truth.

Whoever can change three things: the website, the content, and the company's relationships with the sites that write about it. In a large company those are three teams, and the work stalls in the gaps between them. In a small one they are the same tired person, and the work stalls for lack of hours.

What helps in both cases is making each piece of work small and specific: one passage on one page for one question, one note to one roundup author, one robots.txt line. A specific task gets done. "Improve our AI visibility" does not. The question-by-question method in this guide exists partly for that reason: it turns a vague goal into a list of named pages and dated actions.

It is also the reason Attensira exists. Reading the answers is the easy part, and the workspace does it on a daily schedule. The doing is what an agency charges a retainer for, and what Attensira's agents draft for a person to approve.

In which sources they trust and how often they link. In our own data, Perplexity leaned on reference sites, course pages and directories for definitional questions, while ChatGPT leaned on specialist articles and Google's own pages; for "best GEO tools", the two engines' top six sources did not overlap at all. Our own citation rate was higher on ChatGPT, while our name appeared slightly more often on Perplexity.

The practical rule is to never average engines. Plan, act and measure per engine, because the page that wins one answer may be invisible to the other engine.

The whole guide, for each question you track. Attensira is an AI-native SEO agency run by software: the service of an SEO agency at the price of a tool. This guide's four conditions are the crew's checklist for a lost answer: the workspace reads the answers on each engine on a daily schedule, the Researcher ranks the pages that decide each lost one, and the crew does the work this guide describes: the web engineer fixes access, the writer and copywriter draft the pages and passages, outreach drafts the pitches to the roundups and directories that decide the answer, and the analyst re-reads the answers and reports what moved with its sample size. Workspaces start in Draft, so nothing leaves the workspace until you approve it in one inbox. The win this answer feature shows the deciding sources and the one next action per engine for a single question.

"We just need to rank in Google and the rest follows." It helps, and for Google's AI features it is the floor. It is not enough for engines that run their own indexes and lean on roundups, as our own table shows.

"Outreach to roundups feels like begging." It is closer to correcting the record. Roundup authors want accurate, current information, and most lists are updated. A short, factual note with what changed is useful to them.

"We cannot control what an AI says." True, and nobody else can either. What you can control is whether the engine can read your pages, whether they answer the question plainly, and whether the pages it trusts describe you accurately. Measurement tells you whether those changes moved the answers.

"This will take too long to matter." It takes weeks, not days, because engines recrawl on their own schedules. That is a reason to start with the cheap, certain steps, and to measure from the first day so you can tell when it starts to move.