Skip to content

We hit 92% of client goals last quarter. Schedule a call.

SpikeROAS performance marketing agency logo

AI & Search

How to get cited by ChatGPT: a 2026 GEO playbook

By the SpikeROAS deskPublished Updated 14 min read

The short answer

To get cited by ChatGPT, Gemini or Perplexity you need three things. The right bots must be able to read your page. Your paragraphs must make sense on their own. And someone other than you must back up what you say. The research here is more exact than the advice. One study tested nine edits: adding quotes lifted a page's share of the answer by 41%, adding stats by 33%, and citing sources by 27%. Keyword stuffing made it 9% worse. Google, meanwhile, says plainly that schema markup is not required for AI search, and that it ignores llms.txt files.

What the research actually measured

Almost every GEO article claims to know what works. One peer-reviewed paper actually tested it. Aggarwal and colleagues published GEO: Generative Engine Optimization at KDD 2024. They ran nine content edits against a set of real queries and measured what changed.

Which content changes actually earned more AI citations, per the KDD 2024 GEO studyHorizontal bars showing relative change in position-adjusted word count: adding quotations plus 41 percent, adding statistics plus 33 percent, citing sources plus 27 percent, unique words plus 6 percent, and keyword stuffing minus 9 percent.Change in AI-answer visibility vs. an unedited pageAdd quotations+41%Add statistics+33%Cite sources+27%Add unique words+6%Keyword stuffing9%Source: Aggarwal et al., “GEO: Generative Engine Optimization,” KDD 2024 (arXiv:2311.09735), Table 1.
Relative change in position-adjusted word count — the paper's measure of how much of a generative answer a source occupies. From Table 1 of Aggarwal et al., KDD 2024 (arXiv:2311.09735).
  • Adding quotations from credible sources: +41%.
  • Adding statistics in place of vague description: +33%.
  • Citing sources inline: +27%.
  • Keyword stuffing: −9%. The classic SEO reflex actively hurt.

The paper puts the top three methods at a 30–40% lift, and says the effect changes by topic. This is a lab study, not a promise. But it is the only controlled test we have, and it points the same way three times: be specific and you get quoted.

What Google actually says (and what vendors sell)

Google published its own guide to AI search in May 2026. Most GEO advice you will read was written before it, or argues with it, or both.

Common GEO adviceWhat Google documents
Add schema markup so AI can parse you“Structured data isn't required for generative AI search, and there's no special schema.org markup you need to add.”
Publish an llms.txt file“You don't need to create new machine readable files, AI text files, markup, or Markdown.” Google Search ignores them.
Chunk content into tiny AI-friendly blocks“There's no requirement to break your content into tiny pieces for AI to better understand it.”
Build mentions across the web“Seeking inauthentic ‘mentions' across the web isn't as helpful as it might seem.”
Rewrite your pages in a special AI voiceYou do not need to write in a specific way just for generative AI search.
Left column: advice commonly sold as GEO. Right column: Google's documented position, quoted from its AI features optimization guide.

How retrieval actually works

When you ask an assistant something current, it usually does not answer from memory. It runs a search, grabs a few pages, and writes a summary with links.

How an AI assistant builds an answer, and where you can influence itFive stages: someone asks a question, the assistant searches live, it fetches crawlable pages, it lifts a self-contained passage, and it cites a source. The fetch and lift stages are marked as the two you control.Someone asks“best X for Y”Assistant searcheslive, not memoryIt fetches pagescrawlable HTML onlyIt lifts a passageself-contained winsIt cites a nameyours, or a rival’sThe only two steps you controlCitations are won at retrieval and at the paragraph — not at the ranking.
The two steps you can influence: whether your page can be fetched at all, and whether a paragraph on it survives being lifted out on its own.

Step 1: Let the right crawlers in

This is the boring part that quietly blocks a lot of sites. Most guides also get the bot names wrong.

AgentWhat it doesBlock it and…
GPTBotOpenAI trainingNo effect on ChatGPT search answers
OAI-SearchBotBuilds the ChatGPT search indexOpenAI: opted-out sites “will not be shown in ChatGPT search answers”
ChatGPT-UserFetches a page when a user's prompt needs itLive fetches fail
PerplexityBotIndexes for Perplexity resultsYou stop appearing in Perplexity
Perplexity-UserUser-triggered fetchLive fetches fail
Google-ExtendedGemini apps and Vertex AI trainingGoogle: “does not impact a site's inclusion in Google Search”
GooglebotGoogle Search — including AI Overviews and AI ModeYou disappear from Google entirely

That last pair is the one people get backwards. You cannot opt out of Google's AI Overviews by blocking Google-Extended. AI Overviews and AI Mode run on normal Google Search indexing. Google-Extended only covers Gemini app and Vertex training.

# Search and answer fetchers — allow these if you want citationsUser-agent: OAI-SearchBotUser-agent: ChatGPT-UserUser-agent: PerplexityBotUser-agent: Perplexity-UserUser-agent: Claude-SearchBotUser-agent: Claude-UserAllow: / # Training crawlers — a separate business decisionUser-agent: GPTBotUser-agent: Google-ExtendedUser-agent: ClaudeBotAllow: /
The allowlist we run on this site. Training and search are separate decisions — decide them separately.

One more thing. Google runs JavaScript. Most AI bots do not. If your answer only appears after a script runs in the browser, assume OpenAI and Perplexity never see it.

Step 2: Write passages that survive being lifted

This is the part that moves fastest, and it matches what the study measured. A quotable paragraph has four traits.

  1. 01Self-contained. It makes sense with no paragraph before it. No it, this, or the above.
  2. 02Names the subject. Say the brand, product, or city by name instead of a pronoun.
  3. 03Carries a number and a source. This is the single strongest signal in the research.
  4. 04Short. Search Engine Land's November 2025 audit of ChatGPT-cited posts found 72.4% contained an answer capsule of roughly 120–150 characters — about 20–25 words. We write to 40–60 words as a house rule, which is looser than that finding.
WeakLiftable
Email usually costs less than other channels.Email marketing returns an average of $36 for every $1 spent, according to Litmus.
We help brands improve rankings.SpikeROAS moved a US marketing agency from average position 51.1 to 13.2 in Google Search Console over a nine-month program. Individual results vary.
There are several factors involved.Three factors set solar cost per lead: state incentive levels, roof-eligibility rate, and appointment show rate.
AI search is changing everything.Pew Research found users clicked a traditional search result on 8% of visits with an AI summary present, versus 15% without.
Structured data helps AI understand you.Google states that structured data is not required for generative AI search and that no special schema.org markup is needed.
Same claim, two ways. Only one can be lifted into an answer and attributed.

Step 3: Why the published numbers disagree

Read three GEO articles and you get three answers to one question: do AI citations come from pages that already rank? The numbers look like they fight each other. They do not.

FindingWhen measuredWhat it says
~76% of AI Overview citations also ranked top 10July 2025Citations tracked rankings closely
37.9% of AI Overview citations also ranked top 10March 2026 · Ahrefs, 863K SERPs, 4M URLsThe overlap roughly halved in eight months
In the same March 2026 dataset, 31.2% of cited URLs ranked 11–100 and 31.0% ranked beyond position 100.

The two studies do not disagree. They are eight months apart, and the thing they measured changed. What it means for you: ranking top 10 still helps, but it is not enough on its own, and it is not required either. About a third of citations now come from pages nowhere near page one.

Step 4: Get corroborated somewhere else

Assistants prefer claims they can see in more than one place. If your site is the only page that says what you do, you are a one-source claim.

  • Consistent naming everywhere — same company name, same service names, same city list.
  • Be genuinely present where your buyers already are: industry bodies, review platforms, forums. Do work worth mentioning rather than seeding mentions — Google calls inauthentic mention-building out by name.
  • Original data. Search Engine Land's audit found 52.2% of ChatGPT-cited posts contained original data or branded insight. A small survey or an anonymized benchmark from your own accounts beats another opinion post.
  • Named authors with real credentials attached to the work.

Step 5: Measure it — starting with the free tools

Two free reports shipped in 2026. Most GEO articles still tell you to buy a tool without mentioning either one.

ReportShippedWhat it gives youWhat it cannot tell you
Bing Webmaster Tools — AI PerformanceFebruary 2026Page-level citation counts and the grounding queries behind them, across Copilot and Bing AI answersNothing about Google, ChatGPT, or Perplexity
Search Console — Search generative AI reportsJune 3, 2026Impressions in AI Overviews and AI Mode, by page, country, device and dateNo clicks, no CTR, no queries. Staged rollout

Bing's report is more useful today, because it names the page and the query. Start there. Then fill the gap by hand.

The share-of-model protocol

  1. 01Write 20–50 buying questions in the way people actually ask them. Cover three intents: discovery, comparison, and problem-framing.
  2. 02Run each question 10 or more times per engine. One run is a screenshot, not a measurement — the same prompt returns different brand sets.
  3. 03Log three levels per run: mentioned, recommended, and linked. They are not the same thing and they convert differently.
  4. 04Share of model = mentions ÷ total runs. It is a rate, not a count.
  5. 05Record who appears when you do not. Their cited page is your content brief.
  6. 06Re-run monthly. Compare rates, never single screenshots.

Be honest about the cost. 30 questions × 10 runs × 3 engines is 900 prompts. That is half a day a month, or a script. Anyone selling a one-click version is selling you fewer runs.

Step 6: Attribute it, or it will not survive a budget review

This is where most GEO programs quietly die. The traffic is real, but nobody can point at it.

  • In GA4, check whether `chatgpt.com` and `perplexity.ai` are landing in referral — a lot of assistant traffic arrives with no referrer and lands in Direct instead.
  • Add a “How did you hear about us?” field with an explicit AI assistant option. It is crude and it is the most reliable signal most teams will get.
  • Segment by landing page. AI-sourced visitors tend to arrive deep, on the exact page that answered the question — not on your homepage.
  • Watch assisted conversions, not last click. Being in the answer happens early in the decision.

What still comes from classic SEO

Do not throw out the old playbook. Google says its AI features are built on the same Search ranking systems as everything else. Strong organic work is still the input.

The structure, on-page and internal-link work that moves rankings is the same work that makes a page worth quoting. We wrote up how that played out on a client property, with the caveats, in how to improve average position in Search Console.

Start here this week

  1. 01Open robots.txt. Confirm OAI-SearchBot, ChatGPT-User, PerplexityBot and Perplexity-User are allowed. Decide training crawlers separately.
  2. 02Confirm your answer is in the HTML, not injected by JavaScript.
  3. 03Take your top money page. Put a direct answer at the top, before the story.
  4. 04Rewrite three paragraphs so each survives being copied out alone. Add one number and one source to each.
  5. 05Add a visible FAQ section with five real customer questions. Skip FAQPage markup if Google is your target — it stopped producing rich results in May 2026.
  6. 06Set up Bing Webmaster Tools and open the AI Performance report. It is free and it names your cited pages.
  7. 07Run 10 buying questions × 10 runs on one engine. Write down the rate. That is your baseline.

Questions we get asked

What is generative engine optimization (GEO)?

GEO is the practice of getting your brand retrieved, quoted and linked inside AI-generated answers from tools like ChatGPT, Gemini and Perplexity. It combines crawlability for the right agents, self-contained quotable passages, specific sourced numbers, and corroboration on sites other than your own.

Does structured data help me get cited by AI?

Not according to Google. Its AI features optimization guide states that structured data is not required for generative AI search and that there is no special schema.org markup to add. Keep Article, Organization and Breadcrumb markup for rich results in classic Search, but do not expect schema to buy AI citations.

Do I need an llms.txt file?

No. Google states you do not need to create new machine-readable files, AI text files or Markdown to appear in Google Search, and that Google Search ignores them — adding one neither helps nor harms. Google representatives have compared the idea to the old keywords meta tag.

Do I need to unblock GPTBot to be cited in ChatGPT?

No. GPTBot is OpenAI's training crawler. The agents that matter for citations are OAI-SearchBot, which builds the ChatGPT search index, and ChatGPT-User, which fetches pages during a conversation. OpenAI states that sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers.

Can I opt out of Google's AI Overviews by blocking Google-Extended?

No. Google documents that Google-Extended does not affect a site's inclusion in Google Search, and AI Overviews and AI Mode run off normal Google Search indexing. Google-Extended only governs Gemini apps and Vertex AI training.

Do AI citations only come from pages that rank in the top 10?

Not any more. An Ahrefs study of 863,000 SERPs and 4 million AI Overview URLs in March 2026 found only 37.9% of cited URLs also ranked in the top 10, down from roughly 76% in July 2025. About 31% ranked between positions 11 and 100, and about 31% ranked beyond position 100.

How do I measure AI visibility for free?

Bing Webmaster Tools shipped an AI Performance report in February 2026 that shows page-level citation counts and the grounding queries behind them for Copilot and Bing AI answers. Google Search Console added Search generative AI performance reports on June 3, 2026, but those show impressions only — no clicks, CTR or queries.

What is share of model and how do I calculate it?

Share of model is the rate at which your brand appears across a fixed set of buying questions: mentions divided by total runs. Run each question at least ten times per engine, because a single run is noise, and log whether you were mentioned, recommended, or linked — those three convert very differently.

Does getting cited by AI actually send traffic?

Far less than people assume. Pew Research found users clicked a link inside an AI summary on just 1% of visits, and clicked any traditional result on 8% of visits where a summary appeared versus 15% without one. Treat citations as brand presence at the shortlist stage rather than as a traffic channel.

How long does GEO take to work?

Crawl and format fixes can change citations within a few weeks, because retrieval happens live rather than on an index refresh. Corroboration work — original data, presence on trusted sources, consistent naming — usually takes one to two quarters to move share of model meaningfully.

Sources

Next step

Want your share of model measured?

Send us 10 buying questions. We will run them ten times each across the engines we can reach and show you the rate — yours and your rivals'. The full baseline is 30 questions.

Keep reading

Next step

Get the Spike Brief.

Tell us the number that has to move. We will tell you where the next dollar works — and what we would cut.