Dan Petrovic asked Gemini one million questions and mapped which brands it recommends.
His agency sees Reddit dropped by the models more than 90% of the time, even when Reddit is in the results they are given. He does not track prompts. He says prompt volume does not exist.
Dan runs DEJAN, an Australian agency that has worked on search since 2008. He was the top-rated speaker at the Ahrefs Evolve conference in Singapore, and Tim Soulo interviewed him for an hour on the Ahrefs Podcast.
I watched the whole interview. I run Distribb, and our customers ask me every week how to get their brand named by ChatGPT and Gemini. Dan's answers are the most concrete I have heard, because he tests the models all day and tells you what he ignores. Here is what he said, in the order that makes it easiest to use.
We analyzed 8 top-ranking guides on getting cited by ChatGPT and Gemini and found that only 2 of 8 define grounding, citation and mention as separate outcomes. None of the 8 report how often a grounded source like Reddit gets left out of the final answer.
The podcast that ChatGPT did not recommend
Tim opened with a test. Before the interview he asked ChatGPT, in an incognito window, for the best podcasts for digital marketers. His own show was missing from the answer. He asked Dan what to do about it.
Dan's first answer was short: get better rankings in Bing. He said straight away that he was simplifying, and the rest of the hour is the long version.
→ Start from one question your buyer would ask, and look at what the model answers today.
The models search an index before they answer
The most common mistake, according to Dan, is to think the models crawl the web. They take text in and put text out. The searching is done by tools around them, and those tools query a classic search index.
He described what happened to Tim's question step by step:
- The prompt was broken into several smaller searches, usually three to six. These are called fan-out queries.
- Each search was run against a search engine, and the top results came back.
- The results were filtered and re-ranked, and a shorter list was handed to the model.
- The model did the last re-ranking itself and wrote the answer. Text reads in order, so there is always a first brand, a second and a third.
Which index depends on the assistant. Gemini uses Google. ChatGPT leans on Bing, though Dan says it clearly pulls from other sources too. Claude appears to use Brave, a search engine most SEOs have never looked at. He is careful to call all three black boxes.
→ If you are not in the search index the assistant uses, you are not in the pool it picks from.
Two ways in: the pages already cited, or your own
Once you know which pages the model reads for your question, there are two ways to get into the answer.
The first is to get onto those pages. Dan calls the research step citation mining: run the prompts, collect the sources the assistants use, and study them. If a "best podcasts" article is feeding the answer, being added to that article is the shortest path.
The second is to build your own page that can be selected. His advice is to study what the chosen pages have in common and make a page that looks like them, reads like them and serves the same intent.
Tim applied it to his own case out loud. His podcast had a single thin page on the company site. He could give it a real home with every episode and guest, write about it on the blog, and publish comparisons with other shows, since assistants like to compare options. We covered that last idea in Versus Pages: Why Ross Hudgens Says They Win AI Search.
→ Outreach gets you into an answer faster. Your own pages keep you there.
Three things every index rewards
Does this mean optimizing for three search engines? Dan does not. He puts his effort into Google, because of its market share in search and in AI answers, and he names three things that work in every index.
- Topical centrality. The search engine should have a clear idea of what your site is about. If you do one thing, most of your pages should be about that thing. A large off-topic section blurs the picture.
- Engagement. Search engines watch how people behave on pages. Dan corrects a common idea here: a good signal is not always a long visit. Someone who lands, submits a form in three seconds and leaves got what they wanted. So did a shopper who checks out in one click.
- Popularity. He calls search a popularity contest. Content that a core audience loves, shares and searches for by name does better than brilliant content nobody can follow.
Links and digital PR come on top as a push. He adds one warning. If your whole plan is to copy what competitors already do, you will always be behind them.
→ Keep the site about one thing, make pages that finish the visitor's job, and give people reasons to look for you by name.
What the million questions showed
So far this is search. Dan's research is about what happens after the search.
He describes AI search as a layer that sits on top of the results and interprets them for the user. The model can read results one to ten and recommend number nine first. That makes the model, as he sees it, the main influencer in any buying decision.
Its choice depends on two things. The first is what the model learned about your brand during training: what you sell, what you are connected to, how people talk about you. He calls this the primary bias. The second is what it sees in the search results at the moment of the question.
To measure the first one, he asked Gemini a million questions and built a map of about a million brands and how the model ranks them. One finding he shared: Chinese brands were not favored. BYD, the car maker, was outside the top 100.
The consequence is uncomfortable. If the model has a poor opinion of your brand, it can leave you out even when you rank in the results it was given. If it likes you, it can put you first even when you were last in that list.
→ The ranking gets you into the room. The model's opinion of your brand decides where you sit.
Grounding, citation, mention
People use these three words as if they meant the same thing. Dan separates them:
- Grounding is when your page is handed to the model as possible context for the answer.
- Citation is when the model uses that page to support a sentence and attributes it.
- Mention is when your brand name is written in the answer.
A mention can be plain text or a link. He calls the linked mention the holy grail, and says it deserves more attention than counting citations. Your page can support the facts in an answer while a competitor gets named as the product to buy.
→ When you check an answer, look for your name first, then for a link on it.
Why he does not try to change the training data
If the model's opinion is formed in training, the tempting move is to flood the web with pages saying your brand is the best. Dan gives two reasons it fails.
First, the labs no longer train on everything they find. The leading models already know the language, so the money now goes into curated, high-quality data sets.
Second, spam is cheap to filter. He described an experiment anyone can repeat. Put 100 articles from a blog network in one folder and 100 articles from a respected publication in another. Take a new text, attach it to each folder in turn, compress both with gzip and compare the sizes. The compression alone, with no machine learning, tells you which folder the text belongs to, and he says it gets it right almost every time. The idea goes back to Claude Shannon's work on information entropy.
His picture for a single company trying to shift training data is a person pouring a bottle of water into an Olympic pool.
So he works on what he can change. He takes a client's page, makes a small and precise edit, and checks how the model responds. Then he does it again, in an automated loop, and scores the results with Bayesian inference, the statistical method for updating a belief as evidence comes in. He keeps going until he has a version of the page that the model, as he puts it, cannot resist.
→ He treats the model as fixed and the page as the thing to test.
Google grounds everything, ChatGPT picks one source
Tim asked how often the assistants search at all, compared with answering from memory. The two big ones behave differently.
Google grounds every answer. Dan's examples were questions like the capital of Australia. His explanation is that Google was embarrassed by the early AI Overviews mistakes, and that search costs Google almost nothing because it owns the index. Google also uses several sources for one sentence, so a single claim can carry three or four citations.
ChatGPT is more economical. In its web search, the relationship is one to one: one sentence, one source.
This is where the Reddit number comes from. Many people now say that being on Reddit is the whole job. In the API calls his agency runs, Reddit shows up in the grounding results constantly, because it ranks so well, and the models leave it out of the answer more than 90% of the time.
→ Ranking well in the index does not mean the model will use you. Reddit is the proof.
The model only reads pieces of your page
One more mechanic changed how I look at our own articles. Gemini never receives your full page.
For each fan-out query, Google builds what Dan calls an extractive summary. It lifts exact passages from your page, the ones that score best against that query, and joins them with ellipses. It is longer than a normal search snippet, and nothing is rewritten. The model sees only those pieces.
You do not control where Google cuts. What survives the cut is all the model knows about your product when it writes its recommendation. A sentence that names the product, says who it is for and gives the number can stand on its own. A sentence that starts with "it also does this" cannot.
For the writing side of this, our step-by-step guide on how to optimize for AI search goes through page structure in detail.
→ Write every key sentence so it still makes sense when it is lifted out alone.
Where Distribb fits
Dan runs this with a senior team and his own tooling. Most founders and small agencies cannot staff it. Three of the jobs he describes are things Distribb does on autopilot once you connect your site.
- Getting onto the pages the models read. Distribb's agent finds the roundups and "best of" lists that already rank for your buyers, finds each author's contact and drafts a personal pitch to get you added. On Pro, the research and the drafted emails land in your dashboard. On Accelerator, Distribb runs the whole campaign, follow-ups and replies included.
- Topical centrality. It researches keywords for your business, groups them into clusters, skips the ones you already cover, and writes and publishes the articles to your CMS. Your site keeps growing around one subject.
- Watching the answers. The AI visibility tracker scans ChatGPT, Gemini, Perplexity and Google's AI answers for the prompts your buyers ask, and shows when they cite you instead of your competitors.
You also receive backlinks from real businesses through the backlink exchange, which helps with the plain ranking that everything above depends on.
Pro is $97 a month for up to 30 articles. Accelerator is $495 a month for up to 90 articles, with a human reviewing every piece, plus done-for-you outreach campaigns.
No prompt volume: what he measures instead
The last part of the interview is about measurement, and it is where Dan is the most direct. Tim asked about search volume for prompts. Dan's reply: "There is no prompt volume. Forget about it."
His reasons are simple. The chance of two people typing the same prompt the same way is tiny. Most conversations have several turns. And every answer is personalized by location, device and history.
What he tracks is entities. An entity is the main thing a client wants to be known for, such as "SEO podcast". His method:
- List the entities the client should be visible for.
- For each one, ask the assistant a short, structured question: a customer is looking for this, recommend some brands.
- Keep web search switched on, because the index changes over time while the model stays the same.
- Repeat it on a schedule and record four things: the brand's position in the list, how often it appears at all, its share of mentions and its share of citations.
He blends the last two into a share of voice figure and uses that as the headline number.
If you need a volume estimate, he gives one formula he accepts. Take a prompt, collect the fan-out queries it triggers, look up the search volume of each and add them up. Those are real searches that ran against a real index.
To find prompts worth testing, he shows the model a page and asks what someone would type for it to recommend that page. Or he starts from the entity, or from the top queries in Search Console, and expands them into prompts. If you would sooner buy this than build it, we compared the 9 best AI visibility tools in 2026.
→ Track what you want to be known for, and count how often and how high your name appears.
What I'm taking from Dan
Rank first. Every assistant he discussed pulls its candidates from a search index. If you are missing from the results, nothing else on this list applies.
Get on the lists. The fastest way into an answer is to be added to the pages the model already reads. That is outreach, and it is boring, and it works.
Keep the site about one thing. Topical centrality costs nothing. It only asks you to stop publishing off-topic articles.
Write sentences that survive being cut. The model sees extracts. Name the product and give the number in the same sentence.
Measure entities over time. A weekly list of who gets recommended for your category tells you more than a dashboard of a thousand prompts.
FAQ
Who is Dan Petrovic?
Dan Petrovic is the founder of DEJAN, an Australian search agency based in Brisbane that has operated since 2008. He is known for technical research on how Google and language models select and rank content, including a study in which he asked Gemini one million questions to map its brand preferences.
Which search engine does ChatGPT use?
ChatGPT leans on Bing for web results, according to Dan Petrovic, and also pulls from other sources. Gemini uses Google's index and Claude appears to use Brave. None of the companies fully discloses its sources, so he treats all three as black boxes and puts most of his effort into Google.
Does llms.txt help you appear in AI answers?
Not for the main assistants, in Dan Petrovic's view. He has an llms.txt file on his own site because it is quick to add and harmless. He says ChatGPT, Gemini and Claude fetch pages they found through search, so the file plays no real part in discovery.
What is the difference between grounding, a citation and a mention?
Grounding means your page was given to the model as possible context. A citation means the model used it to support a sentence. A mention means your brand name appears in the answer, as plain text or as a link. Dan Petrovic considers a linked mention the most valuable of the three.
Is there search volume for AI prompts?
No. Dan Petrovic argues that almost no two people type the same prompt, that most chats run over several turns and that answers are personalized. His workaround is to collect the fan-out queries a prompt triggers and add up their search volumes, since those are real searches.
Does being on Reddit get you into ChatGPT answers?
Much less than people think. Reddit ranks well, so it often appears in the results handed to a model. In the API calls Dan Petrovic's agency runs, the models leave Reddit out of the final answer more than 90% of the time.
Start with the lists the models already read
Pick one question your buyers ask, see which pages the assistants read to answer it, and get your brand onto those pages. Then keep publishing around your one subject. If you want both running without you, see how the Distribb AI SEO agent handles the outreach, the content and the tracking, and start the free trial.