Skip to content
SwankyForge
SwankyForge
All insights
Article9 min read

What gets a company cited by ChatGPT and Perplexity: evidence and mentions, not markup

What published research says about being cited in AI answers, which tactics have measured effects, which are hygiene, and a checklist for the engineer who has to implement it.

Fewer people click through to websites when an AI answer sits on top of the results. In a Pew Research Center study of 900 adults' real browsing in March 2025, users clicked a traditional search result in 8% of visits when Google showed an AI summary, against 15% when it did not, and clicked a link inside the summary in 1% of visits. Ahrefs measured the click-through rate of the top-ranked page on keywords with AI Overviews: 34.5% lower in March 2025 than a year earlier, and 58% lower in December 2025 than two years earlier, when the study was repeated. Being the source an AI answer cites now matters alongside ranking underneath it.

A market has grown around this under names like GEO and AEO, and much of what it sells is not backed by evidence. This note separates what has been measured from what is assumed, for the marketing lead deciding where to spend, and ends with a checklist for the engineer who has to build it.

The short version for marketing leads and founders

What works, by the evidence available.

  1. Answer the question in the first paragraph, in plain words, with a number where one exists.
  2. Back claims with statistics, quotations and named sources. This is the one tactic tested in a controlled experiment, and it helped.
  3. Keep important pages current, with real updates rather than a new date stamp.
  4. Get mentioned on other sites. Off-site mentions track AI visibility more closely than anything you do on your own pages.
  5. Describe the company the same way everywhere: one name, one category, one sentence.

What is hygiene, not a lever. llms.txt files and schema markup. They are cheap and harmless, and neither has shown a measurable effect on citations.

Cost and risk. Most of the work is editorial time, plus a few days of engineering for the checklist below. The risk is paying for tactics with no evidence behind them, or for fake mentions that cost more in reputation than they return.

What the controlled study found

The paper that introduced the term is GEO: Generative Engine Optimization by Aggarwal et al., presented at KDD 2024. The authors built a benchmark of 10,000 queries, had a generative engine (GPT-3.5 turbo) answer each one from the top five Google results, then edited one source at a time and measured how much of the answer drew on it.

The results, on the paper's main visibility metric (position-adjusted word count, computed from its Table 1):

Edit to the source page Change in visibility
Add quotations from credible sources +41%
Add statistics in place of qualitative claims +31%
Cite sources +27%
Keyword stuffing −8%

The same pattern held when the authors tested on Perplexity: quotations helped most (+22% on the same metric), and keyword stuffing did 10% worse than leaving the page alone.

Two findings matter for smaller companies. First, the gains were largest for pages that ranked low in search. When every source was edited at once, citing sources raised the visibility of the fifth-ranked page by 115.1% and lowered the top-ranked page's by 30.3%. Second, the authors report no significant improvement from making the text more persuasive or authoritative in tone; the engine responded to evidence, not confidence.

The limits. The engine was a 2023-era model summarising five fixed sources, and visibility means share of a generated answer, not traffic or leads. Commercial assistants retrieve differently and change often. The exact percentages will not transfer. The direction, that verifiable facts beat keywords and adjectives, is the part to act on.

Off-site mentions matter more than on-page tricks

Your site is one input among many. Profound analysed 680 million citations from August 2024 to June 2025. Of the citations going to ChatGPT's ten most-cited sources, Wikipedia took 47.9%; of those going to Perplexity's top ten, Reddit took 46.7%; for Google AI Overviews' top ten, Reddit took 21.0%. The concentration is in the top ten only: Wikipedia was 7.8% of all ChatGPT citations. Even so, a large share of what assistants say about a company comes from pages the company does not control.

Ahrefs studied 75,000 brands and found that the number of web pages mentioning a brand correlated with its visibility in AI Overviews at 0.664, branded anchor text at 0.527, branded search volume at 0.392, and backlinks at 0.218. This is a correlation across brands, not an experiment: large brands have more of everything. But it is consistent with how retrieval works. A model asked about a category assembles its answer from pages that discuss the category, and a company that appears on many of them is more likely to appear in the answer.

What this means in practice: write things other people want to cite (original numbers, methods, honest comparisons), contribute to the places where your buyers ask questions, and make sure directories and profiles describe the company accurately. The same sentence about what you do, on your site and everywhere else, makes it easier for a model to connect the mentions to one entity.

Freshness, measured

Ahrefs analysed about 17 million citations across ChatGPT, Perplexity, Gemini, Copilot and AI Overviews. The pages AI assistants cited were on average 1,064 days old, against 1,432 days for pages in organic search results: 25.7% fresher. ChatGPT cited the newest pages; AI Overviews behaved most like organic search.

Note the absolute numbers. The average cited page was almost three years old. Freshness is a tiebreaker, not a requirement to publish weekly. Updating the pages that answer your buyers' main questions, when the facts change, is enough. Changing the date without changing the content is not an update.

llms.txt and schema: hygiene, not levers

Google's documentation on AI features, last updated in December 2025, says there are no additional requirements or special optimisations for appearing in AI Overviews or AI Mode, that you do not need new machine-readable files, AI text files or markup, and that there is no special schema.org structured data to add. Its guide on optimising for generative AI features, updated in July 2026, names llms.txt directly and says Google Search does not use such files, says structured data is not required, and calls optimising for generative AI search still SEO. It also warns that seeking inauthentic mentions across the web is less helpful than it looks.

A direct measurement agrees. In May 2026 Ahrefs tracked 1,885 pages that added JSON-LD schema between August 2025 and March 2026, against 4,000 control pages. Citations changed by +2.2% in ChatGPT and +2.4% in Google AI Mode, both indistinguishable from zero, and by −4.6% in AI Overviews, a small but statistically significant decline. About 53% of AI-cited pages carry schema, but that reflects sites that do the basics well, not a cause.

Keep valid structured data for the rich results it was designed for, and publish an llms.txt if it costs you an hour. Do not pay for either as a citation strategy.

What not to do

  • Keyword stuffing. It lowered visibility in the controlled study, on both engines tested.
  • Generated filler. Pages written to cover every phrasing of a query add no new fact for a model to quote, and they dilute the pages that do.
  • Fake mentions. Planted forum posts, paid listicles and self-written reviews are what Google's guide warns against, and they are visible to the buyers you are trying to reach.
  • Unsourced numbers. A statistic without a source is the opposite of the edit that worked. If a number cannot be traced, drop it.

An engineer's checklist

  • Answer first. Each page that targets a question opens with a 40–80-word answer that stands on its own when quoted, under a heading that states the question or the finding.
  • Sources on the page. Numbers link to their source, with the publication date, in a visible sources section.
  • Real dates. Published and last-updated dates are visible and match the dates in the page markup and the sitemap. Update them only when content changes.
  • Crawlable text. The main content is in the HTML the server returns, not only after client-side JavaScript runs.
  • Deliberate crawler rules. robots.txt allows the search crawlers of the assistants you want to appear in. Some vendors run separate user agents for search and for model training; decide on each instead of blocking or allowing all of them by default.
  • One description of the company. The same name, category and one-sentence description in the site's organisation markup, about page, profiles and directory listings.
  • Valid, matching markup. Structured data passes validation and matches what is visible on the page. Treat it as maintenance, not a project.
  • Measurement. A fixed list of buyer questions, run monthly in each assistant, with the cited sources recorded; plus referral traffic from assistant domains in analytics. Without this, nobody can tell whether any of the above worked.

The same properties, a direct answer, a cited source and a date, are what a retrieval system needs from your internal documents too. If you are building an assistant over your own knowledge base, the RAG copilot page describes how we test that it answers from the right passage.

Sources

Have a process that costs too much?

Tell us about it. The first call is free, and if AI is not the answer we say so.

Discuss your task