Insights · AI search

How do I get my B2B company cited by ChatGPT, Perplexity and Google AI Overviews?

By Ronan Liedmeier, fractional head of growth · Published · Last updated

Get cited by publishing pages that answer the sub-questions buyers ask, state facts directly with sources, and are readable without JavaScript. Then measure with Bing Webmaster Tools, Search Console and a weekly prompt panel.

You get cited by publishing pages that answer the specific sub-questions a buyer's prompt breaks into, state each fact directly with a source and a date, and arrive as plain HTML that a crawler can read without running JavaScript. Then you measure it weekly with the free reports from Bing and Google plus a fixed panel of prompts, because no paid tool sees more than those three do. That rests on Google's and Bing's own documentation, three Ahrefs studies from 2025 and 2026, one peer-reviewed benchmark, and a crawl failure on my own site that I fixed this week.

By Ronan Liedmeier, fractional head of growth. Last updated 20 October 2026.

What is GEO and how is it different from SEO?

Generative engine optimization (GEO) is the work of making a page eligible to be retrieved, used and cited inside an AI-composed answer. SEO is the work of earning a ranked position for a query. The two share an index, since every major engine starts from web search results, but they reward different things at the selection step.

Bing wrote the term into its official webmaster guidelines in February 2026 and defined it as being about eligibility for grounding and reference in AI responses, with the same caveat as SEO: no guarantee of a citation (Bing Webmaster Guidelines, reported by Search Engine Journal). Google avoids the term and says the opposite of what most agencies pitch: "There are no additional requirements to appear in AI Overviews or AI Mode" (Google Search Central).

Here is how the two differ in practice, from the client sites I have worked on this year.

Dimension SEO GEO
Unit of success A ranked URL for a keyword A cited passage inside a composed answer
What the engine retrieves Pages matching the query Pages matching the sub-queries the engine generated from the prompt
What gets lifted Title, snippet, page One passage that answers one sub-question
Page structure that works Keyword in title, H1, intro H2 phrased as the sub-question, answer in the first two sentences
Proof the engine wants Links, authority, freshness A stated fact, a named source, a date, consistent entity names
Technical floor Indexable, fast, mobile Indexable plus readable without JavaScript
Where traffic goes Click to your page Mostly nowhere: 1% click on links inside AI summaries (Pew)
How you measure Search Console, rank trackers Bing AI Performance, Search Console, a prompt panel

The traffic row is the one founders skip. Pew's July 2025 study of 900 US adults and 68,879 Google searches found users clicked a link inside the AI summary on 1% of visits, and clicked any traditional result on 8% of visits with a summary versus 15% without (Pew Research Center). A citation is a brand mention in front of a buyer at the moment they are comparing options. Treat it as that, and treat the click as a bonus.

How do ChatGPT, Perplexity and Google decide what to cite?

Each engine makes three separate decisions: which pages to retrieve, which retrieved passages to use when writing the answer, and which used sources to show as citations. You can influence the first and second. The third is opaque, and the vendors say so.

Bill Hartzer's review of the published documentation makes the point that Google, OpenAI, Perplexity and Microsoft publish eligibility rules and none of them publish a selection ranking (Hartzer). What they do publish is the retrieval mechanism, and that is enough to work with.

Retrieval starts with query fan-out. Google's documentation says AI Overviews and AI Mode may issue several related searches across subtopics and data sources before writing a response (Google Search Central). OpenAI describes the same step for ChatGPT: the model rewrites a prompt into several targeted queries and sends those to search providers. A buyer who types "best outbound tool for a 10-person SaaS sales team in Germany" triggers sub-queries about tool pricing, LinkedIn automation limits, German cold email law and CRM integrations. The page that gets cited answered one of those, and in the Ahrefs data below most such pages did not rank for the original prompt.

Use follows retrieval. The model reads the retrieved passages and picks the ones that state the answer plainly. Microsoft's February 2026 guidelines are the most concrete public statement of what that means. As Search Engine Journal summarised it, "Facts should be stated directly rather than implied," and entity names should be clear and consistent across the page (Search Engine Journal). If your pricing page says "plans start from a level that works for most teams", the model has nothing to lift. If it says "the Team plan is 49 euros per seat per month, billed annually" (an invented example), it does.

Citation follows use, and this is where the research stops. Analytics Insight's October 2026 summary of the published studies lands where I do: cover the sub-questions in full, publish original data with a documented method, and date your facts (Analytics Insight).

What this means for a B2B site: write each H2 as the sub-query, open it with the answer, and put the number, the source and the date in the first paragraph under it. The rest of this article is built that way on purpose.

Do I need to rank on page one of Google to be cited?

No, and the share of citations that come from page one is falling. In March 2026 Ahrefs analysed 863,000 keyword SERPs and 4 million AI Overview citation URLs and found 37.9% of cited URLs also appeared in the first 10 result blocks (Ahrefs, March 2026). Their July 2025 study of the same question had put it near 76%.

Two caveats before you quote that drop. Ahrefs improved its citation parser between the two studies, and Google moved AI Overviews to a new model in January 2026, so the two figures are not a clean time series. Hartzer's comparison lists other vendors landing anywhere from 32% (seoClarity) to 54.5% (BrightEdge) on different query sets and dates (Hartzer). The direction is consistent across all of them: more than half of what Google cites sits outside its own top 10.

Outside Google the overlap is smaller. Ahrefs ran 15,000 long-tail prompts through the assistants in mid-2025 and found about 12% of URLs cited by ChatGPT, Gemini and Copilot ranked in Google's top 10 for the same prompt. Perplexity was the outlier at 28.6%, because it leans hardest on a live search index (Ahrefs, August 2025).

So ranking is a tailwind, and for Perplexity a strong one. It is no longer the gate. Of the remaining 62% of AI Overview citations in the Ahrefs data, roughly half ranked between positions 11 and 100 and roughly half ranked nowhere in the top 100. Those pages got in by answering a fan-out sub-query, which is a different competition from the head term.

For a founder watching Google traffic fall, the practical consequence is that the pages to build next are the ones that answer the second-order questions in your category, where you can be the best answer on the web with one afternoon of work, rather than another attempt at a head term you have lost to G2 and a software review site.

Which content formats get cited most?

Pages that compare named options in a table, guides that define a term and then answer its sub-questions in order, and pages publishing original numbers with a method get cited more than opinion posts and feature pages. That is what the Bing guidance, the Google documentation and the published vendor studies agree on; what they disagree on is the size of the gap, and I am not going to quote a percentage I cannot source to a URL.

Here is the format ranking I use when planning a client's content calendar, with the reason each format works at the retrieval and use steps described above.

Format Why engines lift it B2B example What it needs to work
Comparison with a table A sub-query like "X vs Y pricing" lands on one row "HubSpot vs Pipedrive for a 15-person sales team" Named options, same criteria per column, a date
Definitional guide with sub-question H2s Each H2 matches a fan-out query "What is a fractional head of growth?" Definition in the first two sentences, then the sub-questions
Original data with a method Nobody else has the number, so the engine has to cite you "What 40 DACH LinkedIn campaigns taught me about reply rates" Sample size, date range, how it was collected
How-to with a worked example Procedural prompts fan out into step queries "How to set up a weekly AI citation check" Numbered steps, a named tool, an expected result
FAQ page Questions map one-to-one onto prompts "Cold email in Germany: 12 questions" Two to four sentence answers, no marketing
Opinion post Rarely: there is no fact to lift "Why most GTM advice is wrong" Publish it on LinkedIn, link to the data piece
Feature or product page Rarely, unless it states specs and prices plainly "Pricing" Numbers on the page, not behind a form

The original-data row is the one most B2B companies underuse. You are sitting on reply rates, churn by segment, implementation times and support ticket categories. Anonymised and rounded, those are facts no competitor can publish, which is the one condition under which an engine has to cite you rather than someone bigger.

I saw the same thing in plain SEO before AI answers existed. At a resource-planning SaaS the founder produced planning templates, I did the keyword research and built download pages with the contact button inside the sheet rather than a form in front of it. The site grew to roughly 10,000 monthly organic visits with a conversion-rate uptick before I left. The pages were ungated and specific, which are the same properties that get a page lifted into an answer today.

Does schema markup or llms.txt help me get cited?

Schema does not, in the only controlled test I have seen, and llms.txt has no engine on record reading it. Google's documentation says it plainly: there is no special schema.org structured data you need to add for AI Overviews or AI Mode (Google Search Central).

The test is Ahrefs' May 2026 study. They tracked 1,885 pages that added JSON-LD between August 2025 and March 2026 against 4,000 matched control pages that never did, and checked citations on Google AI Overviews, AI Mode and ChatGPT. Their finding: "Adding schema produced no major uplift in citations on any platform." AI Overviews showed a small decline of 4.6% relative to controls; AI Mode and ChatGPT moved 2.4% and 2.2%, both indistinguishable from zero (Ahrefs, May 2026).

The wider class of "write for the AI" tricks fares the same. C-SEO Bench, a benchmark from Puerto, Gubri, Green, Oh and Yun presented at NeurIPS 2025, tested the published conversational-SEO rewriting methods and found most were ineffective or harmful to a page's position in the answer. Their second finding matters more for anyone planning a strategy around this: "as we increase the number of C-SEO adopters, the overall gains decrease" (C-SEO Bench, arXiv). The authors call it zero-sum. Whatever small edge a formatting trick gives disappears once your competitors run the same tool.

llms.txt is a proposed convention for a markdown file at your site root describing your content for language models. As of this writing, none of Google, OpenAI, Microsoft or Perplexity document reading it. It costs ten minutes and does no harm, so I add one. I do not report it as work done.

What I keep: FAQ and Article schema on content pages, because they still earn rich results in classic Google, and because Google asks that structured data match the visible text, which is a useful discipline. What I cut from proposals: any line item that sells schema as a way to earn citations.

Does my React or Lovable site even get crawled?

Only if the words are in the HTML your server sends, and on a client-rendered React site they are not. Vercel's December 2024 analysis of crawler traffic across its network found that "none of the major AI crawlers currently render JavaScript", naming OpenAI's OAI-SearchBot, ChatGPT-User and GPTBot, Anthropic's ClaudeBot and PerplexityBot. ChatGPT's crawlers fetched JavaScript files on 11.5% of requests and Claude's on 23.8%, but fetched them without executing them (Vercel). Googlebot does render JavaScript, with a delay, which is why a React site can rank in Google and still be invisible to ChatGPT.

I found this on my own site. On 15 October 2026 I fetched the GMS homepage the way a non-JavaScript crawler would, with a plain HTTP request and no rendering. The response contained the head, a script tag and an empty root div. Every word on the page, including the services, the case descriptions and the author bio, lived in the JavaScript bundle. Google had indexed it. ChatGPT had never seen a sentence of it.

The fix took one working day. I added an Insights section whose articles are written as markdown files and converted to static HTML at build time, so the finished pages ship as text in the initial response, with the React app layered on top for navigation. I re-ran the no-JavaScript fetch the same day and the article body, headings and byline came back in the HTML. That is the crawl test every Lovable, Next.js, Vite or Webflow site should pass before any content work starts.

How to run the test yourself, with no tooling beyond a terminal:

  1. Pick three URLs: the homepage, a service page and your best article.
  2. Fetch each with curl, or with your browser's JavaScript disabled, and save the response.
  3. Search the saved HTML for a sentence from the visible page. If it is absent, crawlers that do not render JavaScript cannot read it.
  4. Check robots.txt for rules blocking GPTBot, OAI-SearchBot, PerplexityBot, ClaudeBot or Bingbot. Some security plugins and CDN defaults block them silently.
  5. Check the server log or CDN analytics for those user agents over the last 30 days. Zero visits from OAI-SearchBot on a site that should be cited by ChatGPT is the diagnosis.

If the test fails, the options in order of cost are: a static build step for content pages (what I did), server-side rendering for the whole site, or a prerendering service in front of the client app. On a Lovable project the first option is a prompt and an afternoon.

How do I measure AI citations without a paid tool?

Three free sources cover more than any paid tracker can see, because two of them are first-party. Bing Webmaster Tools gives you citation counts per URL. Google Search Console gives you AI Overview impressions mixed into Web search. A weekly panel of buyer prompts gives you the ChatGPT and Perplexity view, which neither vendor reports.

Bing's AI Performance report went into public preview on 10 February 2026. It shows total citations, the daily average of unique pages cited, a sample of the grounding queries the AI used when retrieving your content, and page-level citation counts for specific URLs across Microsoft Copilot, Bing's AI summaries and unnamed partner integrations (Bing Webmaster Blog). It is the only first-party report I know of that gives a per-page citation count. Bing is careful about what the number means: page-level counts "reflect how often pages are cited, not page importance, ranking, or placement". Read it as a frequency, and read the grounding queries as a free list of the sub-queries your pages already answer.

Google is less generous. Its documentation says sites shown in AI Overviews and AI Mode are counted in the overall search traffic in Search Console under the Web search type, with no separate filter (Google Search Central). What you can do is watch the ratio: impressions holding steady while clicks fall on a query set is the signature of an AI Overview taking the click, and the 8% versus 15% Pew figures above tell you how much to expect.

The prompt panel fills the gap. Mine for a client is 20 to 30 prompts a real buyer would type, run by hand every Monday in ChatGPT, Perplexity and Google AI Mode, logged in a sheet with four columns: cited (yes or no), which URL, which competitor was cited instead, and the sub-question the answer seemed to be built from. It is an hour of manual work a week. After four weeks the sheet shows which pages the engines lift, which competitors own which sub-questions, and which prompts never cite anyone in your category, which is the gap to fill first.

Two rules for the panel. Keep the prompts fixed for at least a quarter so the trend is real. Write them the way a buyer types them, with the company size, the country and the constraint, because fan-out depends on those details.

What does a GEO engagement with me cover?

It covers an audit against the twelve points below, a fix list ordered by what blocks citation outright before what improves it, a content plan built from the sub-queries the engines already use for your category, and a measurement setup you can run yourself after I leave. It does not cover schema as a deliverable, AI-rewriting of existing copy, or any promise of a citation count.

I do this for B2B SaaS and professional services companies whose Google traffic has been falling since AI Overviews rolled out and who need to be the source the engines quote when a buyer asks about their category. It fits inside a GTM sprint when the content base is thin, or inside a fractional head of growth engagement when the site already has the pages and the problem is structure and measurement. The first step is the crawl test and the prompt panel, which I run before proposing anything, so the conversation starts from what the engines can see today.

The audit I run:

# Check Pass condition Why it matters
1 No-JavaScript fetch Body text present in raw HTML on home, service and article pages AI crawlers do not render JS (Vercel)
2 robots.txt and CDN rules GPTBot, OAI-SearchBot, PerplexityBot, ClaudeBot, Bingbot allowed Silent blocks are common defaults
3 Crawler log check Visits from the above user agents in the last 30 days Confirms they can reach you
4 Google index and snippet eligibility Pages indexed, no nosnippet or max-snippet limits on content pages Google's only stated requirement
5 Bing index and Webmaster Tools Site verified, AI Performance report active Only per-URL citation data available
6 H2s as buyer sub-questions Each H2 is a question a buyer would type Matches query fan-out
7 Answer-first sections First two sentences under each H2 answer it What the model lifts
8 Facts stated directly Prices, numbers, dates and names on the page, none implied Bing grounding guidance
9 Consistent entity names One spelling of the company, product and people across the site Bing grounding guidance
10 Sources and dates visible Each data claim links to a source; each page shows an updated date What engines and buyers both check
11 At least one comparison or data page per category A table comparing named options or original numbers with a method Highest-citing formats
12 Measurement running Prompt panel sheet, Bing report, Search Console query set Without it the work is a guess

Points 1 to 5 are blockers. A site that fails any of them is invisible to at least one engine regardless of what the content says, and that is where the first week goes. Points 6 to 11 are the content work, done page by page starting with the category's highest-intent sub-queries. Point 12 is what makes the next quarter's decisions possible.

Re-check by 20 January 2027: the Ahrefs overlap figures, whether Google has added an AI filter to Search Console, whether any engine has documented llms.txt, and whether the Bing AI Performance report has left preview.

FAQ

Do I need to rank on page one of Google to be cited in AI Overviews?

No. Ahrefs' March 2026 study found 37.9% of AI Overview citations also ranked in Google's top 10, so most did not. For ChatGPT, Gemini and Copilot the overlap was about 12% in Ahrefs' August 2025 test. Ranking helps, but the engines pull from pages that answer the sub-questions they generate, and those pages can sit anywhere in the index.

Does adding schema markup help my pages get cited by ChatGPT?

The only controlled test I know of says no. Ahrefs compared 1,885 pages that added JSON-LD against matched controls and found no meaningful uplift on Google AI Overviews, AI Mode or ChatGPT. Google's own documentation says no special structured data is required. Keep schema for rich results, and spend the time on the content instead.

Can ChatGPT read a React or Lovable website?

Only if the text is in the HTML the server sends. Vercel's analysis of its network traffic found that OpenAI, Anthropic and Perplexity crawlers fetch pages but do not execute JavaScript. My own homepage served an empty root div to non-JS fetchers until I moved the content to a static build step. Test with a fetch that has JavaScript turned off.

How do I measure AI citations without paying for a tracking tool?

Three sources. Bing Webmaster Tools has a free AI Performance report with per-URL citation counts across Copilot and Bing. Google Search Console folds AI Overview impressions into the Web report without a separate filter. A weekly panel of 20 to 30 buyer prompts run by hand in ChatGPT, Perplexity and Google covers the rest.

What is the difference between GEO and SEO?

SEO earns a ranked position for a query. GEO earns a citation inside an answer that an engine composed from several pages. The retrieval overlaps, since all engines start from a search index, but the selection step differs: GEO rewards pages that state a fact plainly, with a source and a date, in a passage that can be lifted on its own.

Sources

Related

Re-check by 2027-01-15