GEOforge Research · Citation Study
AI tools don’t read the whole web and pick the best page. They rely on a small group of websites and quote the page that best answers the question. That changes where you should publish, and how.
By Delcho StanimirovGEOforge10 min read
Every week, GEOforge records which websites ChatGPT, Google AI Overviews and Google AI Mode link to when they answer our clients’ buyer questions. So far that’s 3.7 million citations, meaning links to sources inside AI answers. We used them to find out what separates the websites and pages that get cited from the ones that don’t. It comes down to three jobs: know which websites AI already trusts for your topics, get your content onto them, and make every page answer one specific question in the words your buyers use.
A small group. Out of 71,386 websites that AI answers linked to, the top 1% (714 sites) received 58% of all citations. At the other end, 42% of websites were cited only once, and together they add up to less than 1%. Eighty websites appear across ten or more different industries and take 17% of all citations by themselves.
The biggest pattern in the study was about websites, not pages. If a website already shows up in AI answers to a brand’s other buyer questions, it’s far more likely to be cited for the next one. That mattered several times more than anything we could measure about the page itself, and it held for all four brands we tested, in completely different industries.
Less than most marketers expect. Only 2.6% of the links in AI answers point to the brand’s own website, and 1.4% for the typical brand. That’s close to an independent study by Featured, which found brand websites made up 1.7% of 22,881 Perplexity citations, and to the 2% figure we discussed on The GEO Show.
That’s a share of links, not a share of answers. A typical AI answer lists 8 to 18 sources, and your website usually takes one of those spots. Across 22 brands, the share of answers that linked to the brand’s own site ranged from 1% to 49%. Competitors’ websites together take about 29% of citations. The other two thirds go to publications, review sites, forums, video and reference sites that no brand in the category owns.
Whatever you do to your own website, you’re competing for one spot in a list of a dozen or more. Most of that list is written by someone else.
GEOforge checks two things in every AI answer: does it recommend the brand, and does it link to the brand’s website? Here’s how the two compare.
| AI engine | Recommended when it links to you | Recommended when it doesn’t | Recommendations that came with a link to you |
|---|---|---|---|
| ChatGPT | 98.9% | 2.3% | 77.0% |
| Google AI Overviews | 77.1% | 3.0% | 84.8% |
| Google AI Mode | 76.1% | 5.4% | 78.8% |
The pattern is the same on all three. When an answer links to your site, it very likely recommends you. When it doesn’t, it almost never does: only 2% to 5% of the time. Around four in five recommendations come with a link to the brand’s own website.
One word of caution: this shows the two go together, not that one causes the other. An AI that has already decided to recommend you will naturally link to you. Either way, the goal is the same: be one of the sources in the answers that matter to you.
To separate the page from the website, we compared two pages on the same website, offered as sources for the same question, where one was cited 40 times or more and the other three times or fewer. Same website, same reputation, same question, so the difference has to be the page. We read the pages, then repeated the whole test on a second, fresh set.
| The cited page had… | First test | Second test |
|---|---|---|
| More of the question’s words in its headline | 77% | 78% |
| More of the question’s words in its text | 76% | 73% |
| Text more focused on those words | 70% | 66% |
| More words overall (length) | 54% | 52% |
Length made no difference. A coin toss would score 50%, and length scored 52–54%. One page cited 114 times was under 300 words long, and it beat a page four times longer on the same site that was cited once.
Matching the question made a big difference. Across both tests, the cited page was the better match 68% of the time, and the odds of that happening by luck are tiny, about 4 in 100,000. When we also counted words with the same meaning, like “cost” and “pricing” or “lawyer” and “attorney”, the cited page was the better match 71% of the time.
All three AI engines prefer the page that matches the question
How often each engine cited the page that better matched the question’s meaning. 50% would be a coin toss.
Each engine was judged only on its own citations of the same pages: 51 page pairs for ChatGPT, 101 for AI Overviews and 107 for AI Mode.
The AI engines use different sources, but they pick pages the same way. That makes this one piece of GEO advice that works everywhere: write the page that answers the question being asked, in the words people use to ask it.
Yes, but faking it doesn’t work. Adding this year to your copy makes no difference. Real age does.
Keeping important pages up to date isn’t housekeeping. It’s part of staying in AI answers.
Mostly not. Across 730 questions answered by more than one engine, ChatGPT shared only about 31% of its sources with Google’s AI search, while Google’s AI Mode and AI Overviews shared 61% with each other. Of the ten sources each used most for a question, ChatGPT and Google had fewer than three in common. ChatGPT leans on research and institutions, such as PubMed and Nature. Google leans on YouTube, Reddit and Instagram, which ChatGPT almost never cites.
The good news is that a brand’s overall standing carries across. Brands that do well on one engine tend to do well on the others, even though the sources behind them differ. ChatGPT just mentions brands less often: 18% of its answers named the brands we track, compared with 28% for Google. So you need one brand strategy, but a different source strategy for each engine.
Put together, the findings point to a clear process. This is what GEOforge is built to do.
If the page that matches the question wins and length doesn’t, then covering more questions beats writing longer guides. You need a focused page for each question your buyers ask. That takes volume, which is what ContentForge is for. It turns your knowledge base and strategy into topic ideas, links each topic to the question it should answer, and writes drafts at scale. You can add topics one by one or upload them in bulk from a spreadsheet. When a new idea overlaps something you’ve already published, ContentForge suggests updating the existing page instead, which is exactly what the freshness results reward.
The websites that matter are the ones already showing up in answers to your buyers’ questions, not a generic list of high-authority sites. CiteForge builds that list from your own tracking: every page any AI engine linked to for your questions, how often, and whether your competitors are on it and you aren’t.
A website AI cites a lot is only useful if it will publish you. The Scout agent checks each one for a single thing: does this site publish articles written by companies, with a byline? It only gives a yes or no when it can point to a page on that site that proves it, such as contributor guidelines. It skips government and university sites automatically, and flags sites that only take paid sponsored posts, so they aren’t mistaken for free coverage. You get a ranked list of sites to pitch, based on how often AI cites them for your questions, how many of your competitors appear there, and whether they accept contributed articles. The Scout is available on GEOforge’s Dominate and Enterprise plans.
The same rule applies on other people’s websites. AI prefers the page that answers a specific question, so a guest article should focus on one buyer question, not a general thought-leadership piece. CiteForge’s Pitch Artist writes the outreach email using your knowledge base and the target page, so your pitch reflects what you actually know.
We’d rather tell you now than have you find out later. We tested whether being mentioned on someone else’s page makes AI more likely to recommend you. We read 664 of the pages AI was citing and checked how each one talked about the brand. At first, pages that featured a brand seemed linked to many more recommendations. But when we compared similar pages cited for the same question, being mentioned was worth only about three percentage points, which could easily be chance. Our test could reliably spot an effect of about 13 points or more, so a smaller effect may still exist.
In short: these are the right websites to target, and a page that answers the question is the right thing to put there. But no one should promise that a single placement will raise your numbers by a set amount. That’s why off-site work should be judged by whether the prompts it targets improve, not by how many placements you get. GEOforge tracks the same prompts every week, so you can see what happened to them before and after each piece goes live.
Citations. 3.7 million citations recorded by GEOforge on ChatGPT, Google AI Overviews and Google AI Mode for 29 brands across B2B and consumer industries, up to September 2026. Google’s own redirect links were left out. Counts come from our tracking, not from website traffic.
Page comparisons. We downloaded 365 pages in September 2026, in two separate rounds. Each pair was two pages on the same website for the same question: one cited 40 or more times, the other three times or fewer. That gave 52 usable pairs in the first round and 83 in the second, across about 120 questions. Pages that didn’t load or showed a bot check were left out, and that happened about equally to cited and uncited pages.
Other figures. The engine comparison covers 730 questions answered by more than one engine. Own-website figures cover 22 brands. The test of mentions on other people’s websites covered 664 pages and 84,762 AI answers, mostly from two brands, so treat it as a first result rather than a final answer.
GEOforge shows every source AI cites for your buyer questions, finds the websites worth pitching, and writes content that answers the questions your buyers actually ask. 14-day free trial, no sales call.
Start a free trial →Sources. GEOforge citation data (3.7 million citations, 29 brands, three AI engines); 365 pages compared in September 2026; 664 cited pages reviewed for brand mentions. Outside comparison: Featured’s analysis of 22,881 Perplexity citations (August 2026).
Delcho Stanimirov is Head of Product at GEOforge, where he leads how the platform measures and improves brand visibility in AI search. A former head of paid media and analytics, he works on making AI visibility as measurable and accountable as paid search.
