ChatGPT Is Secretly Googling Things: This Tool Shows You Exactly What

When your customers inquire ChatGPT aliases Gemini something, the exemplary softly fires a group of accepted web searches successful the background, retrieves the ranking pages, and synthesizes the reply from those. The sites that rank for those hidden queries get cited. The ones that don’t, don’t. QueryFan generates persona-specific prompts, runs them done some models, and captures the nonstop searches each 1 triggered. That database is your existent AI visibility target. It’s free.

Keywords Lists Are Useful, They Just Miss Half The Picture

Let maine beryllium precise astir that earlier anyone writes a furious reply.

I’m utilizing the word “keywords” to mention to the “one-shot” queries that spell into accepted hunt engines. Yes, I cognize we’ve been successful a “semantic” world for complete a decade, but let’s conscionable work together connected terminology that everyone tin travel for now.

The superior rumor of “keyword lists” successful discourse to AI hunt is threefold:

1. Typically, queries (prompts) that spell into LLMs thin to beryllium longer, multifaceted, and conversational successful nature. Traditional searches thin to beryllium much constrictive successful scope.
2. Traditional hunt is “one-shot.” You do your search, get your information, past do different independent search. Queries/prompts connected LLMs thin to beryllium conversational successful quality and transportation the discourse of erstwhile tokens.
3. The mechanisms that LLMs usage for web hunt besides transportation personalization context. If the personification has antecedently stated they are a vegan, and they inquire the LLM astir [running shoes], it is highly apt the LLM will execute a hunt to accommodate this.

In essence, AI hunt has go a benignant of “universal intent decoder” for users. Those big, multifacted conversations pinch the AI get surgery down into subsets of solvable queries, which are tally successful the inheritance arsenic “traditional” searches connected Google aliases Bing, pinch the resulting sites utilized to make a response. The process is known arsenic “Retrieval Augmented Generation” (RAG).

Many users are unaware that “traditional” searches are happening successful the inheritance (Image Credit: Mark Williams-Cook)

The optimization target has moved. You are nary longer optimizing purely for what the quality types into a chat box. You are optimizing for what the AI supplier softly searches for connected their behalf, successful the background, without the personification knowing it happened.

Those inheritance queries are what QueryFan captures. They are often rather different from what the personification really asked. And they are the nonstop database of things you request to rank for to look successful AI-generated answers.

Exhibit A: Reddit Fell Off A Cliff On A Tuesday

The scope and extent of this concealed narration became clear when Reddit was enjoying meteoric visibility increases successful Google, and calamity struck connected September 10th, 2026. According to citation search information from PromptWatch, Reddit’s citation complaint successful ChatGPT responses collapsed almost overnight. It had been moving arsenic precocious arsenic 15% of each citations. Within days, it was sitting beneath 2%.

The origin was unglamorous: Google softly removed the expertise to petition 100 hunt results simultaneously (the num=100 parameter) from its hunt API connected that date.

Reddit’s citations successful ChatGPT collapsed erstwhile Google removed num=100 (Image Credit: Mark Williams-Cook)

Think astir what this tells you. Reddit’s visibility successful ChatGPT responses tracked Google’s bulk hunt capabilities, not thing Reddit did, not a training information update, not an alignment tweak. The accusation is astir arsenic subtle arsenic a dropped piano: ChatGPT was bulk-pulling Google hunt results, Reddit dominated those results astatine the time, and erstwhile the bulk-pull disappeared, truthful did Reddit’s citations.

AI hunt surfaces are, successful ample part, wrappers astir accepted search. The “AI” spot is existent (the synthesis, the personalisation, the conversational coherence) but the information retrieval measurement is remarkably familiar. Google indexes and ranks the web; the AI consults that index. Your contented still needs to rank.

How QueryFan Works

An overview of QueryFan.com logic (Image Credit: Mark Williams-Cook)

Step 1: Your ‘Traditional’ Keywords

Your accepted keyword database for the word “running shoes” whitethorn incorporated various suggested variations of this term, from a root for illustration Google Suggest.

For QueryFan.com, we tin simply return the overarching taxable (Image Credit: Mark Williams-Cook)

For QueryFan, we tin simply return the taxable of “running shoes” and usage this arsenic our first step, arsenic we are going to make prompts astir this.

The first QueryFan measurement to participate the taxable (Image Credit: Mark Williams-Cook)

Step 2: Define Personas

Your personas are really we are going to customize the prompts we generate. This will change our traversal of the token space, aligning america pinch training information from the millions of communities, forum posts, Reddit threads, and net sermon wherever existent users inquire existent questions pinch these identities.

QueryFan sends your persona + taxable operation to the LLM to make the kinds of questions that persona would really inquire an AI tool. Not keywords. Questions. Real, conversational, context-laden questions. For the [middle-aged vegan man who conscionable started running] example, it will nutrient things like:

• “Which vegan moving shoes are bully for middle-aged men conscionable starting to run?”
• “Where tin I bargain vegan moving shoes online successful the UK?”
• “What should I look for erstwhile choosing my first brace of moving shoes arsenic a beginner?”

Step 3: LLM Selection And AlsoAsked Enrichment

AI conversations branch. Someone who asks astir vegan moving shoes will inquire follow-up questions: astir cost, astir brands, astir wounded prevention. QueryFan passes the generated prompts done the AlsoAsked API to seizure the nearest-intent follow-up questions astir each one. People Also Ask information is the correct instrumentality present because it was built to exemplary mobility proximity, which is precisely what you request erstwhile you’re trying to foretell wherever a speech goes next.

For instance, a hunt successful the UK for “running shoes” would aboveground travel up questions connected circumstantial brands, asking really to prime a shoe, and moreover communal aesculapian queries.

AlsoAsked mobility character for “running shoes” showing nearest intent proximity questions (Image Credit: Mark Williams-Cook)

You tin besides prime if you wish to usage ChatGPT, Gemini, aliases both. Each LLM handles and instrumentality retired queries somewhat differently, truthful if you’re optimising for a circumstantial level it is champion to get the information from there.

QueryFan configuration surface (Image Credit: Mark Williams-Cook)

Step 4: Query Fan-Out

QueryFan sends the enriched punctual database to GPT-5 pinch web hunt enabled (via the OpenAI Responses API) and to Gemini pinch Google Search grounding progressive (via the Gemini Grounding API). Both models, erstwhile they determine a punctual requires existent information, execute existent Google searches down the scenes.

This process captures the fan-out queries arsenic some APIs are, alternatively usefully, transparent astir what they searched. The Gemini API returns a webSearchQueries array successful the groundingMetadata section of each grounded response. OpenAI’s Responses API logs the existent hunt queries successful the web_search_call output. QueryFan harvests both.

The consequence is simply a table: persona-specific prompts in, the existent Google hunt queries the AI fired out. Not what your customer typed. What the AI searched for connected their behalf. Those are your caller SEO targets, and until now location has been nary free instrumentality that surfaces them astatine scale.

The Grounding Question: Not Every Prompt Triggers A Search

A little but important caveat earlier you sprint disconnected to categorize everything arsenic an SEO opportunity.

Not each punctual causes the AI to execute a web search. The models make a determination based connected the statement of token prediction arsenic to if unrecorded accusation is required.

To springiness an example, the punctual “What do reddish humor cells do?” doesn’t trigger a search. The logic is location is simply a very steep bell-curve of which tokens are going to look next. In the billions of training documents, the reply has stayed very stable, truthful an “in-model” reply tin confidently beryllium generated.

At the other extremity of the scale, a punctual specified arsenic “What happened successful the news today?” would trigger a web search. There would beryllium a very level curve of “wtf tokens are next?,” arsenic location is nary “stable” reply wrong the training data; it ever changes, it requires unrecorded data. It’s different type of the Query Deserves Freshness (QDF) conception that SEOs person utilized for years.

If you’re willing successful grounding, Dan Petrovic has done immoderate fantabulous activity successful this area, and moreover released trained models connected Hugging Face to foretell whether queries will beryllium grounded erstwhile they deed a assurance threshold.

In-model answers are very slow to alteration (Image Credit: Mark Williams-Cook)

QueryFan surfaces which prompts triggered searches and which didn’t. Only the grounded ones (the ones that really caused a Google hunt to happen) are actionable done SEO. The in-model answers are, for now, mostly extracurricular your reach. You’d request to power training information to move the needle there, which is simply a different task entirely, pinch a overmuch longer horizon.

What You Do With The Results

You now person a database of existent hunt queries that AI devices occurrence erstwhile answering questions from your circumstantial personas. Run a modular spread analysis:

• Which of these queries do you person contented for?
• Which do you already rank for?
• Which person zero coverage, either connected your tract aliases anyplace you’re apt to beryllium mentioned?

The first 2 categories are diagnostic. The 3rd is your action list.

Example results from QueryFan.com (Image Credit: Mark Williams-Cook)

One important favoritism from accepted SEO: Your own ranking isn’t the only way to AI visibility. LLMs scan the apical 10, 20, sometimes 50 results for a grounded query and synthesize crossed them. A trusted reappraisal tract ranking astatine position 3 is simply a morganatic way to appearing successful an AI-generated answer, moreover if your ain domain ne'er makes the first page. Getting a merchandise reviewed connected a high-authority master site, earning a mention successful a roundup article, appearing successful applicable organization content, each of these count.

LLM visibility is simply a multi-site focus. This intends the spread study has 2 outputs: contented to create on your ain site, and placements to gain on different people’s sites.

The Punchline

Cast your mind backmost to that Reddit citation graph. The 1 that fell disconnected a cliff erstwhile Google changed a azygous API parameter. An wholly independent company’s AI visibility tracked the behaviour of a hunt API it didn’t power and astir apt didn’t cognize existed.

That’s the style of the dependency. And the accusation isn’t that SEO is dead; it’s almost the opposite. SEO is now operating astatine 1 further remove: alternatively of optimizing for the quality query, you request to optimize for the AI-translated query that happens betwixt the quality and Google.

QueryFan gives you a measurement to spot what that translator really produces. Your keyword database tells you what group typed into a hunt bar. QueryFan tells you what ChatGPT and Gemini searched for connected their behalf, successful the background, without anyone asking them to denote it.

Those are different lists. The spread betwixt them is not a insignificant refinement to your contented strategy. It’s the portion of AI hunt that cipher has been measuring because cipher has had a free instrumentality to measurement it with.

Disclosure: The writer is the creator of Queryfan.

More Resources:

• • New Data Finds Gap Between Google Rankings And LLM Citations
  • Google’s New AI Search Guide Calls AEO And GEO ‘Still SEO’
  • The AI Convergence Problem

This station was primitively published connected Mark Williams-Cook Substack.

Featured Image: Roman Samborskyi/Shutterstock