GEO (Generative Engine Optimization) encompasses practices aimed at making a website or brand appear in the responses of ChatGPT, Gemini, Perplexity, Copilot, or Google’s AI Overviews. It is first based on SEO fundamentals: pages accessible to crawlers, indexed, useful, and verifiable, as well as the brand’s presence in reliable third-party sources. No assistant publisher discloses its selection criteria: this guide distinguishes what is documented, what is observed, and what remains hypothetical.
The topic has gained importance because conversational assistants have become a gateway to information, including for commercial queries. But it also carries many unverifiable promises. The data cited below was verified in November 2026; tools and platform rules evolve quickly and must be rechecked before any decision.
What is Generative Engine Optimization?
The term comes from a research paper titled “GEO: Generative Engine Optimization”, posted on arXiv in November 2023 by a team including Princeton University and IIT Delhi, and later accepted at the KDD 2024 conference. The authors define “generative engines” as systems that search for sources on the web and then write a synthesized response citing them. GEO refers to methods that increase a content’s visibility in these responses. To learn more, read “What is GEO? Definition and Stakes.”
The vocabulary is not settled. One also encounters AEO (Answer Engine Optimization), LLMO, or simply “AI SEO.” Google, in its optimization guide for its generative AI features published in May 2026, considers these terms cover the same work and that, for its search engine, optimizing for generative search remains SEO. The core question, what GEO is compared to classic SEO, is not just a change of label: it’s mainly the way of measuring results that changes, and that’s where the essential comparison SEO vs GEO lies. We detail this comparison in “SEO vs GEO: Differences, Commonalities, and Complementarities.”
Four Results to Never Confuse
Being “visible in an AI” can refer to very different situations, which have neither the same value nor the same measurement tools. Mixing them leads to misleading dashboards.
| Result | Definition | Example |
|---|---|---|
| Citation | Your page appears as a linked source of the response | A numbered link to your article under a Perplexity response |
| Mention | Your brand is named in the text, with or without a link | ChatGPT cites your company among three providers, without a link |
| Referral Traffic | A user clicks and lands on your site | A session with an assistant as the referrer in your analytics tool |
| Conversion | That visit leads to a useful action | Quote request, sign-up, purchase |
A brand can be often mentioned without ever being cited, or cited without receiving clicks. Conversely, a decision made in a conversation can later result in a direct visit or a brand search, which tools will attribute to another channel. This is the main limitation of everything that follows.
How Assistants Search for and Cite Sources
An assistant can respond in two ways. Either it relies solely on what the model learned during its training, with no displayed source; or it launches searches at the time of the question, retrieves pages, and then writes a response citing them. This second method, called retrieval-augmented generation or grounding, is the only one on which a site can act quickly, as model training is done on data frozen at a given date.
Google describes this operation for its AI Overviews and its AI Mode: its usual ranking systems retrieve pages from the search index, and the model can launch several related queries in parallel, a technique Google calls query fan-out. According to its documentation for publishers, a page simply needs to be indexed and eligible for display with a snippet to serve as a reference link, with no additional technical requirement. Optimizing for AI Overviews therefore first comes down to succeeding at Google SEO. We detail this point in “Google AI Overviews: How to Prepare Your Site for AI Responses?”
Who Crawls What: The Bots to Know
Each publisher uses its own bots, with distinct roles. Confusing them is the most costly technical error: blocking a search bot thinking you’re blocking model training removes the site from responses. This topic is covered in a dedicated article: “AI Bots and robots.txt: Who to Allow and How to Check?”
| Publisher | Bot or Token | Documented Role | Effect of a Block in robots.txt |
|---|---|---|---|
| OpenAI | OAI-SearchBot | Make sites appear in ChatGPT search | The site is no longer displayed in ChatGPT search responses |
| OpenAI | GPTBot | Collect content that can be used for training | Signals that content should not be used for training |
| OpenAI | ChatGPT-User | Visits triggered by a user request | robots.txt rules may not apply |
| Perplexity | PerplexityBot | Make and link sites in results | Respected according to Perplexity; the site loses its chance of being cited |
| Perplexity | Perplexity-User | Retrieve a page at a user’s request | Generally ignores robots.txt, according to Perplexity |
| Googlebot | Crawling for search, including AI features | The site drops out of search and thus AI Overviews and AI Mode | |
| Google-Extended (robots.txt token) | Training of Gemini models and grounding in Gemini applications | No effect on inclusion or ranking in Google Search |
These roles are drawn from OpenAI’s bot documentation, Perplexity’s, and the list of Google’s bots, consulted in November 2026. OpenAI and Perplexity indicate that a robots.txt modification may take about 24 hours to be taken into account. Google also announces its intention to support, in early 2027, the standardized training preference signal AIPREF. The details of AI bot and robots.txt configuration deserve a line-by-line check, as a firewall or CDN can also block these bots without the file being the cause.
What the Foundational Study Measured, and What It Doesn’t Prove
The 2023 research paper built a benchmark of 10,000 queries (GEO-bench) and tested nine content modifications on a simulated generative engine, then on Perplexity. The best methods, adding source citations, textual citations, and statistics, improved measured visibility by about 30 to 40% in relative terms, according to the authors. Keyword stuffing brought no gain and even degraded results in the Perplexity test.
Another result is often cited: sources ranked fifth in search results gained 115.1% visibility with added citations, while top-ranked sources lost an average of 30.3%. These figures describe a controlled experiment, with models available in late 2023 and a metric defined by the researchers. They indicate a plausible trend—substantiated and precise content is more reused—but do not constitute a rule of operation for current assistants.
Which Technical and Editorial Fundamentals to Master?
The starting point is often lower than one thinks. The Fasterize barometer published in November 2026, based on a crawl of 1,469 sites, noted that 61% of sites scored 80/100 on its Search pillar versus 9% on its AI Search pillar: a gap detailed in our article on poor site optimization for AI search. These scores are based on the vendor’s grid, but they show that basics remain the first area of work.
On the technical side, four conditions are documented. The assistants’ search bots must be able to access pages, including through the firewall and CDN. Pages must be indexed and eligible for snippets, as Google specifies that the nosnippet, data-nosnippet, and max-snippet directives also limit what appears in its AI features. Important content must exist as text, as Google requires; and since assistant publishers do not document their bots’ ability to execute JavaScript, it’s better for it to be directly in the served HTML. Finally, structured data must match the visible text. To go further, read our article “Structured Data and AI: Does Schema.org Change Visibility?”
Several misconceptions fall away when reading official documentation. Google indicates it does not use the llms.txt file, that no specific schema.org markup is needed for its AI features, and that splitting pages into small blocks is not useful. The debate on llms.txt and AI SEO remains open for other services, but no major player has, to our knowledge, documented using it to choose sources. As for schema.org, it retains its value for rich results and information consistency, without being a free pass. To learn more, read “llms.txt: Does This File Really Help AI SEO?”
Since August 31, 2026, Google also offers in Search Console a “Search generative AI” setting that allows excluding a site from AI Overviews, AI Mode, and generative features in Discover. This setting is set to “include” by default; exclusion removes all traffic and impressions from these features, with no effect on the rest of search according to Google. Check this in each property, especially after a site migration or takeover.
Control Grid and Level of Evidence
The following table classifies the most cited levers based on the strength of supporting evidence. “Documented” means a publisher writes it in its documentation; “observed” means a study or test noted it under specific conditions; “hypothesis” means it’s a reasonable practice without published proof.
| Lever | Level of Evidence | How to Check |
|---|---|---|
| Search Bot Access (OAI-SearchBot, PerplexityBot, Googlebot, Bingbot) | Documented | robots.txt, server logs, firewall rules |
| Indexing and Snippet Eligibility | Documented (Google) | URL Inspection in Search Console |
| “Search generative AI” Setting to “include” | Documented (Google) | Search Console Settings |
| Key Information Present in Served HTML | Prudent Hypothesis (text documented by Google, JavaScript execution not documented for others) | Compare source code and rendered page |
| Precise Facts, Sourced Figures, Citations | Observed (GEO Study 2023) | Review key pages: does each claim have a source? |
| Clear Titles, Tables, FAQ | Recommended by Microsoft for Bing and Copilot | Editorial review of most strategic pages |
| Presence in Third-Party Sources (directories, comparisons, press) | Observed in Ad Hoc Tests | List sources cited for your key queries |
| llms.txt File | Hypothesis; Ignored by Google Search | No Public Indicator |
How to Optimize a Brand’s and Its Content’s Visibility?
On your own pages, convergent publisher recommendations boil down to a few principles. Google asks for “non-interchangeable” content, produced from real experience or expertise, rather than syntheses of what already exists. In the announcement of its AI Performance report, Microsoft advises structuring information (headings, tables, FAQ), supporting claims with examples and data, regularly updating pages, and signaling changes via IndexNow.
Concretely, a page is more likely to be reused if it explicitly answers a question, provides verifiable elements (starting price, timeframe, scope, update date, method), and indicates who is speaking and in what capacity. These markers overlap the “useful content” criteria Google has described for years: who produced the content, how, and for what purpose.
A brand’s visibility is not only played out on its own site. Assistants also rely on third-party pages: sector directories, comparisons, professional profiles, press articles, reviews. Hence the importance of the brand entity, i.e., consistent information (name, activity, address, offerings, executives) everywhere the company is described. For businesses and products, Google explicitly cites Merchant Center and the Google Business Profile; Microsoft, for its part, cites Bing Places. Being cited by ChatGPT follows the same logic, with the particularity of depending on the assistant’s web search and its OAI-SearchBot. We detail this point in “Brand Entity: How to Make a Company Identifiable by AIs?”
Practices That Backfire
- Producing pages en masse to capture assistant queries. Google reminds that creating pages primarily to manipulate its results or generated responses violates its rule against large-scale content production.
- Manufacturing mentions. Google’s guide indicates that seeking “inauthentic” mentions on the web is not as useful as it seems; its anti-spam systems also apply to generative responses.
- Stuffing pages with keywords. This is the only method tested in the 2023 GEO study that brought no benefit.
- Blocking the wrong bot. Blocking OAI-SearchBot or PerplexityBot to “protect content” removes the site from responses, whereas GPTBot and Google-Extended suffice to express a training refusal.
- Believing in guaranteed ranking. Google specifies that no third-party tool has access to its internal ranking or AI systems. Any offer of “guaranteed position” in an assistant should be treated with suspicion.
How to Measure Mentions, Citations, Traffic, and Conversions?
Measurement is the weakest link in GEO, and each platform gives a partial view. Measuring AI visibility therefore requires combining several data sources, knowing what each covers and what it misses.
Since June 2026, Google provides a generative AI performance report in Search Console, available to all sites since August 31, 2026. It counts impressions in AI Overviews and AI Mode (an impression each time a link to the site is displayed), by page, country, date, and device. It does not provide clicks, and this data also appears in the classic performance report: adding both counts them twice.
Since February 10, 2026, in preview, Microsoft offers an AI Performance report in Bing Webmaster Tools. It tracks citations in Copilot, in summaries generated by Bing, and with some partners, with “grounding queries,” the expressions used by the AI to retrieve your pages, presented as a sample. Views by intent, by theme, and a “citation share” were added in June 2026; Microsoft specifies this citation share does not represent a traffic share.
For ChatGPT, Perplexity, and the Gemini app, there is no public equivalent of these reports to our knowledge: Perplexity and Gemini SEO is tracked via referral traffic analysis and query testing. According to OpenAI’s publisher FAQ, links from ChatGPT search carry the parameter utm_source=chatgpt.com, making them easier to spot in an analytics tool.
| Metric | Data Source | What It Measures | Main Limitation |
|---|---|---|---|
| Google AI Impressions | Search Console, Generative AI Report | Display of a link in AI Overviews or AI Mode | No clicks; AI Overviews and AI Mode not separated |
| Copilot and Bing Citations | Bing Webmaster Tools, AI Performance | Number of citations, pages, and retrieval queries | Preview; neither clicks nor traffic |
| Referral Traffic | Analytics Tool (referrers, UTM parameters) | Visits from an identifiable assistant | Some visits arrive without a referrer |
| Mentions Without Links | Hand-tested query panel or third-party tool | Presence of brand name and wording | Variable responses, limited sample |
| Conversions | Analytics, CRM, “how did you hear about us?” question | Business value of the channel | Incomplete journey attribution |
Volumes remain modest at the site scale. According to Similarweb data cited by Fasterize on 498 sites in August 2026, assistants sent a median of 2 visits per 100 visits from organic search. Several analyses, however, note higher conversion rates for these visitors, on limited samples: this is a reason to track conversion as much as volume.
A Reproducible Manual Testing Protocol
- Write 20 to 50 questions your customers actually ask, spread across informational, comparative, and commercial, without mentioning your brand.
- Choose the assistants to monitor and note for each the mode used (web search enabled or not), language, and account (logged in or not).
- Ask each question several times, in separate sessions, as responses vary from one try to another.
- For each response, note: brand mentioned or not, pages cited, competitors present, wording used, any factual errors.
- Archive dated screenshots, then repeat the protocol at fixed intervals (e.g., every month) with the same questions.
This protocol does not give market share, but a trend on a chosen sample. Commercial GEO tools automate this work at a larger scale; their method (number of queries, frequency, location, models queried) must be known before comparing their figures.
GEO Audit Plan and Mistakes to Avoid
A GEO audit doesn’t need exotic tools. It consists of checking, in order, that nothing blocks access, that content is usable, that the brand is correctly described elsewhere, then establishing a baseline measurement.
- Access. Read robots.txt and firewall or CDN rules for OAI-SearchBot, PerplexityBot, Googlebot, and Bingbot; check server logs that these bots get 200 responses.
- Training Choice. Decide separately on GPTBot and Google-Extended, which concern training and, for Google-Extended, grounding in Gemini applications.
- Indexing. Check in Search Console and Bing Webmaster Tools the indexing of strategic pages and the absence of unintended nosnippet directives.
- Google Setting. Verify the Search Console property is set to “include” in the “Search generative AI” setting.
- Rendering. Ensure prices, features, and key answers are in the served HTML, not only after JavaScript execution.
- Content. Identify pages that answer key questions and add dated facts, sources, and methodological elements.
- Entity. List listings and profiles describing the company and correct outdated or contradictory information.
- Baseline Measurement. Enable Google and Bing reports, isolate assistant referrers in the analytics tool, and run the first round of the testing protocol.
Judging a Result Without Drawing the Wrong Conclusion
An increase in citations after a change does not prove the change caused it: models change, competition publishes, demand varies. To limit this bias, compare a group of modified pages to a group of comparable pages left unchanged, over the same period, and keep a dated record of each change. Microsoft itself reminds that citation activity depends on model evolution, competitor content, and demand.
Finally, set proportionate goals. For a content site, the expected gain is mainly a presence in cited sources; for a business, it’s being mentioned correctly when a buyer asks for a recommendation. In both cases, GEO extends SEO and reputation work rather than replacing it, and no method guarantees a spot in an assistant’s response.
