AI search optimization is the practice of making your brand the one AI engines name and cite when they answer a buyer's question. GEO (generative engine optimization) names it from the research side, as getting into answers that generative engines compose. AEO (answer engine optimization) names it from the search side, as being the single answer returned. LLM SEO names it from the model side, as being what ChatGPT or Claude says about you.
To do this in GetIntel, see Get recommended in AI search.
Disclosure: GetIntel sells AI visibility tracking, so we have a commercial interest in this topic. Every figure below is our own measurement or a quote from the company it describes, with the date and a link to the underlying data, and where our evidence is thin we say so.
GEO vs AEO vs LLM SEO: how the terms compare
The five terms in common use describe one practice from different starting points. None of them has a formal definition that practitioners agree on, so the useful question is what each one emphasises.
| Term | What it emphasises | Who coined or uses it | Where it overlaps with SEO |
|---|---|---|---|
| GEO (generative engine optimization) | Being included and cited in answers that engines compose from several sources | Named in a 2023 research paper led by Princeton researchers; now the usual term among agencies and AI visibility vendors | Crawlable pages, clear entities and authority signals |
| AEO (answer engine optimization) | Being the one answer returned, from featured snippets and voice assistants through to AI answers | No single coiner; SEO practitioners from the featured-snippet and voice-search era, before LLMs | Question-led headings, direct answers, structured data |
| LLM SEO | What a language model says about you, including answers drawn from training data rather than live search | Informal; founders and marketers talking about ChatGPT and Claude specifically | Least overlap: training data is not something a crawl or ranking fixes |
| AI SEO | Ambiguous: used both for SEO done with AI tools and for optimizing for AI answers | SEO suites and content platforms | Heavy, because half its usage is ordinary SEO |
| LLM visibility | The measurement rather than the practice: how often engines name or cite you | Tracking vendors, as the name of the metric | None directly; it measures an outcome SEO tools do not report |
In practice you do not pick one. The work that gets you cited is the same across all of them, and the outcome they share has its own name, AI findability: whether AI crawlers can reach you, answers name you and answers cite you. The distinction that changes what you do is between engines, not between acronyms, which is what the next section is about.
How do AI engines actually pick sources?
Each engine picks sources differently, and they barely agree with each other. The figures below come from two measurements of our own. The first put 100 buyer questions across 10 mainstream software categories to five engines on 2 August 2026, producing 491 answers and 3,754 citations. The second is GetIntel's daily tracking of 130 buyer questions in the AI visibility category across four engines from 10 July to 10 August 2026.
Across the first run, the 3,754 citations went to 1,258 different domains. The most-cited domain took 3.5% and the top ten together only 14.8%, while 57% of domains were cited exactly once, as covered in no top-10 list of AI citation sources. Between engines, the overlap is small: on the same questions, Gemini and ChatGPT shared 4.5% of their cited domains, and the closest pair, Google AI Overviews and Gemini, shared 12.7%. Being cited in one engine tells you little about another.
One thing holds across all of them: naming a brand and citing a source are different acts. An engine can recommend you by name with no link to your site, or cite your page while recommending a competitor. The difference between being mentioned and being cited decides what you should count.
ChatGPT
ChatGPT names the most brands and cites the fewest sources. On 2 August 2026 it named 5.28 vendors per answer, the most of the five engines, while citing 3.66 domains per answer, and it returned 387 citations across 100 answers, the fewest of the five in total. 27 of its 100 answers cited nothing at all while still recommending products by name. Its sources are also the least stable: across 1,050 consecutive daily run pairs, ChatGPT's cited source set changed 99.7% of the time, with an average overlap of 16.7% between one day and the next. Measuring it through the API is not the same as measuring the real interface: in 11 paired tests on 11 August 2026 the two shared 13.6% of their sources, detailed in API vs the real ChatGPT interface. For why a competitor gets named instead of you, see why ChatGPT cites competitors instead of your brand.
Perplexity
Perplexity is the most consistent engine and the most willing to recommend nobody. It cited a median of 9 domains per answer on 2 August 2026, but the distribution is split: 84 of 100 answers cited six or more sources, 16 cited none, and none cited between one and five, as counted in our Perplexity source study. It named no vendor at all on 19 of 100 commercial questions. On our AI visibility question set on 17 September 2026, it named no product on 71.7% of 120 observations. It is also the steadiest: its cited set changed on 76.4% of consecutive daily runs, against 99.7% for ChatGPT, with a day-to-day overlap of 51.4%. The practical audit is in Perplexity citation sources data.
Gemini
Gemini cites few sources and changes them constantly. On 2 August 2026 it cited 3.5 domains per answer and named 4.53 vendors, and it was the only engine that never answered a buying question without naming a vendor. Its source set changed on 99.0% of consecutive daily runs, with the lowest day-to-day overlap of any engine at 10.1%. On our AI visibility questions on 17 September 2026 it named no product on only 5% of prompts. Because it is grounded in Google Search, ranking gets a page considered, but citation is a stricter test of whether a clean passage can be lifted, which Gemini source selection walks through.
Claude
Claude spreads its citations more widely than any other engine. On 2 August 2026 it cited 5.27 domains per answer across 403 distinct domains in 100 answers, and its top three domains took only 3.6% of its citations, against 19.6% for Google AI Overviews. It named 4.22 vendors per answer and named none on 3 of 100 questions. We track Claude weekly rather than daily, so we do not have a day-to-day churn figure for it. The engine-specific guide is how to rank in Claude.
Google AI Overviews
AI Overviews cites the most sources and concentrates them the most. On 2 August 2026 it cited 10.05 domains per answer, never answered without citing, and named the fewest vendors at 3.59 per answer. It is also the most concentrated engine: its top three domains took 19.6% of its citations, and YouTube alone took 11%. It does not always appear: on 5 August 2026 it rendered no Overview for 25 of 100 questions, while Google AI Mode answered all 100 and cited 16.4 domains per answer. The two Google surfaces shared only 21.8% of their cited domains, set out in AI Mode vs AI Overviews vs Gemini, and the patterns within Overviews are in Google AI Overviews citation patterns.
Who gets cited: vendors or third parties?
The split between vendor and third-party citations depends on the category's age. In the AI visibility category, measured on 6 August 2026 across 9,049 citations from the 40 most-cited domains, 51.7% pointed at companies selling a tool in that category and 5.7% at independent editorial, as covered in half of AI's tool advice comes from tool vendors. Reddit, YouTube, LinkedIn, Medium and Instagram together took 26.0%. The most-cited single pages were vendor roundups: Zapier's best AI visibility tools page carried 154 citations on its own. In a young category, engines cite what has been written, and most of what has been written is vendor marketing.
What's the same as SEO and what isn't
Most of the foundation is shared. What changes is where the work lands and what you count.
| Same as SEO | Different from SEO |
|---|---|
| Crawlability: engines cannot cite what their crawlers cannot fetch or render | Mentions on third-party sites matter more than backlinks: engines read the pages that talk about you, linked or not |
| Clear entities: one consistent description of what your product is and who it serves | Being in listicles and Reddit threads matters, because those are what engines quote for buying questions |
| Authority signals: a domain engines already have reason to trust | Structure for extraction is table stakes, not an edge: pages with question-led headings made up 65.1% of pages cited five or more times and 65.8% of pages cited once, on 6 August 2026 |
| Fresh, accurate content that answers the question | llms.txt makes no measurable difference, on our data and on Google's own statement (below) |
| Measuring on a fixed query set over time | You count mentions and citations per engine, not positions, because position in an answer does not persist |
On technical access, the most-cited pages were not doing anything special. Of the 39 most-cited domains in our AI visibility tracking on 18 August 2026, none blocked a named AI crawler in robots.txt, 35 of the 37 we could fetch were readable without JavaScript, and 29 had an llms.txt file. Having the file is common among cited sites, but it is not why they are cited: across the 50 most-cited domains we found no measurable effect from llms.txt, and across the 40 most-cited pages on 11 August 2026, pages with JSON-LD averaged 118.3 citations against 121.3 for pages without it. Google's guidance on AI features in Search, last updated 10 December 2025, says there are "no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary," and "You don't need to create new machine readable files, AI text files, or markup to appear in these features." Both are cheap and harmless; neither is the lever.
Position is the other real difference. In our daily tracking from 10 July to 10 August 2026, a domain cited in two consecutive runs held the same position only 39.5% of the time, and the first-cited source changed in 67.7% of run pairs. A rank tracker makes sense because search positions persist. In an AI answer they do not.
SEO is not dead either: Google ranking and AI citation behave as separate machines. We compared two competitors directly: llmrefs.com ranks for 967 Google keywords and earns 0 AI citations, while otterly.ai ranks for 67 and earns 969. Run them as two programmes with different inputs.
AEO vs SEO: what's the difference?
AEO aims to be the answer an engine gives, while SEO aims to rank a page in a list of links. SEO is judged by position and clicks; AEO by whether your brand is the one named when an engine answers a question. The groundwork overlaps heavily, because a crawlable, well-structured page that answers a question clearly helps both. The measurement does not: a page can rank first on Google and never be quoted in an answer, and a page that ranks nowhere can be cited because it answers one specific question better than anything else. AEO also predates large language models. It began with featured snippets and voice assistants, where there was only ever one answer to win. If you are choosing software for this, our comparison of the best answer engine optimization tools ranks them on what engines actually cite.
A practical order of operations
The sequence below is the one we would follow for a founder starting from nothing. Each step depends on the one before it.
- Measure where you stand. Put your real buyer questions to the engines and record who is named and which sources are cited. The free AI citation checker asks 20 buyer questions on ChatGPT, Gemini and Google AI Overviews in about two minutes, with no signup, and gives you a baseline to beat.
- Find who is cited instead of you. The gap is rarely your own site. It is usually a roundup, a review profile or a Reddit thread that names a competitor, and finding the Reddit and G2 sources AI cites for competitors shows how to trace them. To see which rivals ChatGPT names, see how to find the competitors ChatGPT recommends.
- Fix the pages AI cannot extract from. Lead each section with the answer, keep your product description consistent across pages, and put comparisons in tables. The GEO content scorer checks a page for the obvious problems.
- Get listed where AI looks. Review sites, directories, honest comparison roundups and the communities your buyers use. Our list of 234 SaaS directories is a starting point, and getting cited without an agency covers the free placements.
- Track weekly, per engine. Because sources change on nearly every run, one check is a snapshot. Track the same questions on each engine separately: ChatGPT brand tracking, Perplexity brand tracking, Gemini brand tracking, Claude brand tracking and the AI Overviews tracker.
- Iterate on the questions that moved. Judge a change over several weeks of repeated runs, not one before-and-after pair. How many reruns an AI visibility reading needs gives the numbers.
Common mistakes
These come from what our own tracking data shows, not from a survey of customers.
Optimising for one engine. Engines share between 4.5% and 12.7% of their sources on the same questions, measured on 11 August 2026. Work aimed at ChatGPT's sources can leave you invisible on Perplexity, and a single blended score hides which engine moved.
Chasing domain rating instead of placement. We have not measured DR against citation directly, and we will not claim a correlation we have not tested. What we have measured is where citations land: in the AI visibility category, the single most-cited pages were specific roundups, led by one Zapier page with 154 citations. Getting onto the pages engines already cite moves more than raising your own domain's authority.
Publishing volume instead of answers. In our 2 August 2026 run, 57% of cited domains were cited exactly once. Engines reach for the page that answers the exact question, not the site with the most pages on the topic. Ten thin posts on one theme compete with each other; one page that answers the question completely is what gets lifted.
Treating AI visibility scores as precise. Answers are sampled from a system that changes between runs. When we split 320 tracked series into halves with nothing changed in between, 13.1% still swung by 10 points or more. A score read once, or compared across two readings, can report an improvement that did not happen. Why AI visibility scores can be misleading covers the causes.
Treating it as a one-off project. The sources engines favour change continuously, so a citation won in August is not guaranteed in October. This is ongoing maintenance, closer to keeping a product listing current than to shipping a site migration.
If you would rather compare tools than build the measurement yourself, our rankings of the best AI visibility tools and of AI citation tracking tools are measured against what engines actually cite. Definitions for every term used here are in the glossary.
Last updated 26 September 2026. Changelog: 26 September 2026, merged four definition posts (GEO, AEO, LLM SEO and AI search engine optimization) into this page and added per-engine source data from our own tracking.