ChatGPT Already Knows Who It’ll Recommend Before It Searches

I requested ChatGPT for the perfect AI note-taking app. Seven phrases, no model names.

Earlier than it fetched something, it wrote itself this search.

finest AI word taking apps 2026 official pricing options Granola Notion AI Otter Fireflies Fathom Mem Limitless

Learn the tail of that string. Granola. Notion AI. Otter. Fireflies. Fathom. Mem. Limitless. Seven merchandise in a single search. I’d named none of them, and nothing had come again from the online search but, so no consequence put them there. These names got here out of the mannequin.

Then it ran 9 extra searches.

web site:granola.ai pricing options AI assembly notes 2026
web site:otter.ai pricing AI assembly notes 2026
web site:fathom.video pricing AI assembly assistant 2026
web site:notion.com product AI Assembly Notes official 2026 pricing
web site:fireflies.ai pricing official AI assembly notes 2026
web site:mem.ai pricing AI notes official 2026
web site:notebooklm.google official options pricing 2026

That’s the fan-out I wrote about within the first two elements, and now you may see what it’s fabricated from. One search to call the shortlist, then one search per identify on that shortlist, each pointed straight at that company’s own website.

So the fan-out was by no means a seek for candidates; it was ChatGPT taking place a listing it already had, one identify at a time.

Within the first two elements, I wrote that you must survive a web site:yourdomain.com probe, as a result of ChatGPT runs them. It solely runs one on you when you had been named in that first search. Miss the shortlist and your web site by no means will get checked out, nonetheless properly constructed it’s. The choice occurs earlier than something touches your server.

I’ve spent two months studying this visitors for part 1 and part 2. That is the very first thing I’ve discovered that adjustments what I’d inform a consumer to do.

Earlier than you learn any additional, all of this comes off one account, so each proportion under is a path somewhat than a measurement. The mechanism is a distinct matter. You possibly can reproduce that by yourself account in two minutes, and I’ll present you ways on the finish.

The 4 Concepts Behind This

Every one units up the subsequent.

1. ChatGPT writes its personal search queries, and you’ll learn them.
2. These queries already comprise manufacturers no one talked about.
3. Being in that question is value about 33 occasions greater than being findable.
4. When you’re in, a second and far harsher filter decides who will get cited.

Concept 1: It Writes Its Personal Searches, And They’re Readable

Once you ask a query, ChatGPT rewrites it into search queries of its own, runs them, reads what comes again, then writes a solution.

These queries sit within the response your browser downloads, underneath a key presently referred to as search_queries. OpenAI renamed it from search_model_queries in early August 2026. Nothing about this can be a leak. Your personal browser wants that JSON to attract the web page, and you’ll learn it in DevTools by yourself account in about two minutes. Recipe’s on the finish.

Right here’s one from a query about reside chat software program.

finest AI reside chat buyer help software program 2026
Intercom Fin Zendesk AI pricing official

I requested for “finest ai based mostly reside chat help software program.” It added the 12 months, “official,” “pricing,” and three names.

All the things under comes from studying a couple of hundred of those.

Begin right here. Open ChatGPT, ask the “finest [your category]” query your consumers ask, and skim the question it writes again. That one string reveals you whether or not ChatGPT is aware of your model exists, which is what AI visibility audits are offered to reply. All the things else on this article is about what to do with what you discover in it.

Concept 2: The Shortlist Exists Earlier than The Search Runs

The apparent objection to my note-taking instance is timing. Possibly ChatGPT searched as soon as, noticed these manufacturers, then wrote a better second question. That’d make the names a results of retrieval somewhat than a trigger.

So I examined it correctly. For every dialog, I took the primary person message and the primary search question, ordered by timestamp. At that time nothing’s been fetched, so there’s no earlier consequence to have realized from.

In 21 of 27 conversations, that first question contained manufacturers the person by no means typed.

Have a look at rows two and three. Similar query, barely completely different phrasing, and the checklist grew from three names to seven. The checklist stretches with the way you ask as a substitute of popping out of a hard and fast desk.

Then I ran  12 classes that don’t have anything to do with one another, to test this wasn’t a software program factor.

11 of 13 did the identical.

The robotic vacuum row is the one which received me. Recalling that Roborock exists can be unremarkable. It recalled Saros, Dreame X50, and Eufy S1 Professional. Present mannequin numbers, within the first question, unprompted. No matter this data is, it goes right down to the product line.

The electrical SUV row does one thing completely different from all of the others. For accounting and remedy and internet hosting, it named distributors and went to their pricing pages. For vehicles, it named magazines. Automobile and Driver, Edmunds, Prime Gear. In a single class, its intuition is to go to the makers; in one other, the reviewers.

I received enthusiastic about that after which re-ran three classes to see if any of it holds nonetheless.

Language studying barely moved. 5 of six names got here again the identical. Accounting collapsed from six distributors to a single focused probe at QuickBooks. Internet hosting switched sides fully, dropped each vendor and went to a assessment web site as a substitute.

So distributors versus magazines is a bent that may flip between runs, not a hard and fast property of the class. Two issues survive that. The injection occurred each single time, and the class with the clearest market leaders stored its names. My guess is that settled classes have steady shortlists and contested ones wobble, however three repeats isn’t sufficient to say that.

Don’t decide your AI visibility from one reply. Run the query 5 occasions, as a result of the checklist adjustments between runs.

It Occurs Each time ChatGPT Has To Provide The Merchandise

Each question up so far had the phrase “finest” in it, in order that was the very first thing I attempted to interrupt. I ran twenty-four extra queries that averted the phrase, throughout seven completely different shapes.

“Greatest” turned out to don’t have anything to do with it. What issues is whether or not ChatGPT has to provide you with the merchandise itself.

When it doesn’t search in any respect, none of this is applicable. Ask how noise-cancelling works, or what a vector database is, and it solutions from coaching with no net search. Similar for open-ended moaning. “We’re spending an excessive amount of on buyer help tooling” received no search both. Seven of my twenty-four by no means touched the online, so there was no shortlist to get onto.

Identify the manufacturers your self, and it takes them. “Xero or QuickBooks for a small enterprise” went straight to web site:xero.com/uk. “Ought to I exploit HubSpot for a small company” went to web site:hubspot.com.

Go away the candidates to it, and it reaches into reminiscence. That occurred in ten of the eleven circumstances the place I requested for a suggestion with out naming something.

The third row is my favourite. I intentionally averted saying “robotic vacuum” and named no one. It went to a particular Roborock mannequin’s personal web site on the primary search.

Row two ought to fear you when you promote software program. That’s a grievance, not a purchasing query. Any individual grumbling about their assembly notes received became Granola official pricing earlier than a single web page was fetched.

And there’s a model of this aimed straight at your rivals. All three of my displacement queries pulled in new names. “Alternate options to Zendesk” produced Assist Scout; “what can i take advantage of as a substitute of quickbooks” produced Zoho Books; and “one thing like duolingo however higher for grammar” produced Kwiziq and Babbel. When your buyer goes on the lookout for a means out of your rival, ChatGPT nominates whoever it already is aware of. That may very well be you, and also you don’t get a say in it on the day.

If ChatGPT searches for a product and also you haven’t named one, it brings its personal.

Do that. Run your class query 5 occasions and write down the names that seem within the question every time. The names that present up each run are your actual aggressive set in ChatGPT’s head. Something that comes and goes is contested floor, which is the place a push can transfer one thing. In case your model by no means seems in 5 runs, you might have your reply, and it isn’t a technical one.

Concept 3: Being In The Question Is Value About 33 Instances Extra

I sorted each model into two teams. Ones that appeared in a question ChatGPT wrote, and ones that had been solely fetched throughout the search with out ever being named. Then I checked how usually every group made it into the ultimate reply.

About 33 occasions the distinction.

I additionally discovered 86 circumstances the place a model received beneficial, and its web site was by no means fetched in that dialog in any respect. A point out doesn’t want a crawl.

That is the uncomfortable half for my very own business. Most of what’s offered as GEO proper now could be retrieval work, which is the two.1% column. The 68.9% column is set earlier than any of it runs.

Do that. Break up your price range to match the 2 columns. When you’re absent from the question, the cash belongs within the work that will get you written about, reviewed, in contrast, and listed. Digital PR, class content material, review-site placement, analyst protection, exhibiting up within the roundups your consumers learn. When you’re already within the question, that spend is essentially accomplished, and the technical work under is what’s left.

Concept 4: Being In The Question Is The Entry Ticket, Not The Win

If it stopped there, the recommendation can be “construct model fairness,” and we may all go dwelling. A second filter runs after the question, and it’s brutal.

I constructed a labelled dataset from 57 conversations. Each retrieved web page as a row, with whether or not it earned a quotation because the label. 3,554 pages.

110 received cited. That’s 3.1%.

ChatGPT reads round 600 pages to write down one reply and credit about 30. Virtually all the pieces will get learn. The hole between being learn and being credited is the place the work is.

Three issues separated the cited from the ignored.

Place, laborious. ChatGPT teams outcomes by area, and the place you sit in that group predicts almost all the pieces.

Beneath the highest two, quotation is a rounding error. Being within the retrieved set isn’t a win when you’re ninth.

Piling on pages hurts. When a number of pages from one area present up in the identical group, per-page conversion collapses.

Two tightly matched pages is the candy spot. Previous six, you’re largely competing with your self. I’ve been telling shoppers to consolidate for a 12 months on intuition. That is the primary time I’ve watched it occur within the information.

Relevance qualifies you; it doesn’t choose you. I scored the cited web page in opposition to each different web page retrieved for a similar declare, on how properly its textual content matched the sentence being supported. The cited web page sat within the high 5% of the pool. However it was the only finest match solely 20% of the time, and its common overlap was properly under the perfect accessible.

So declare relevance builds a shortlist and one thing else picks the winner. Get into the highest 10% for a particular declare, and also you’re within the dialog. That half you management by writing. The ultimate choose entails issues I can’t see.

Do that. Take the questions your consumers ask and discover each web page of yours that solutions the identical one. That’s the underside row of the desk above, and it’s costing you. Choose the web page that matches the intent most tightly, make it the reply, and fold or redirect the remainder into it. Then be certain the sentence that really solutions the query sits close to the highest of that web page in plain HTML textual content, with the numbers in it.

Discovering these overlaps by hand throughout an actual web site is the depressing half. Cannibalization is generally judged on key phrase overlap, and that’s the improper unit right here, as a result of ChatGPT teams by declare somewhat than by key phrase.

That is what Key phrase Insights was constructed for.

It clusters your key phrases by search intent somewhat than by string match, so the pages combating over one intent floor as a gaggle even once they share no key phrases.

That grouping maps carefully onto what I’m watching ChatGPT do when it collapses a site down to 1 cited web page.

Full disclosure: It’s my firm, which can also be why I do know it handles this particular job.

What I’d Truly Do With This

The findings cut up AI visibility into two video games that maintain getting handled as one.

Recreation One Is Being In The Class Vocabulary

If ChatGPT doesn’t already join your model to your class, it gained’t identify you within the question, and also you’re a 2% likelihood of a point out.

Schema gained’t repair that, and neither will web page pace. An llms.txt file has even much less of an opportunity, as a result of your server by no means will get contacted earlier than the choice’s made.

What appears to construct it’s sluggish and unglamorous. Being written about, reviewed, in contrast, and argued over throughout the open net for years, till the affiliation exists within the coaching information.

That’s digital PR and category-defining content material, which is awkward for everybody presently promoting technical audits as an AI technique.

Recreation Two Is Profitable The Quotation As soon as You’re In

That’s the three.1%, and it’s actual. One tightly matched web page per intent, the claim-bearing sentence early, info and numbers in plain HTML textual content, and no cluster of near-identical pages combating one another.

Right here’s the entire thing as a sequence you may run this week.

1. Learn the question. Ask your class query 5 occasions and word the manufacturers ChatGPT writes into its personal search. Ten minutes.
2. Discover which recreation you’re in. Within the question each time means you’re taking part in recreation two. By no means in it means recreation one, and no quantity of web page work will transfer you.
3. When you’re absent, go and get written about. Critiques, roundups, comparability pages, the publications your class reads. In a class like electrical SUVs, the place ChatGPT went to the magazines earlier than it went to any producer, that’s the one door.
4. When you’re current, test conversion, not retrieval. Rely how usually your pages get fetched in opposition to how usually they get credited. Fetched quite a bit and credited hardly ever is a web page drawback, not a visibility drawback.
5. Repair the web page, not the positioning. One web page per intent, the reply sentence close to the highest, actual numbers in HTML textual content somewhat than in a picture or loaded by JavaScript.
6. Re-check month-to-month. The steadiness check confirmed the shortlist strikes, and the format strikes with it. In early August 2026, OpenAI renamed the important thing these queries sit underneath, and the fan-outs in my captures dropped from 12 searches per reply to 4. What was true in July gained’t essentially maintain in September.

In my information, one model received fetched 66 occasions and by no means cited as soon as. That’s not an consciousness drawback. The engine stored going again and stored deciding that they had nothing value quoting. Step 4 exists to catch that.

Steps one and 4 are what I constructed FanoutFox for, as a result of operating them by hand for each reply will get previous quick. It’s a free Chrome extension; all the pieces stays in your browser, and it reads your personal ChatGPT session somewhat than guessing from the surface. It reveals you the fan-out queries verbatim, so the first step is a look as a substitute of a DevTools session.

For step 4, it now has a converts column, retrieved in opposition to cited per area, the quantity this text argues you have to be monitoring as a substitute of retrieval counts. A site that will get learn repeatedly and by no means credited is the analysis you’re on the lookout for, and it’s sorted proper there.

The Caveats

I’d somewhat you belief the elements that deserve it.

The mechanism is stable. ChatGPT writing model names into its first question is seen in a single seize, and you’ll reproduce it your self in two minutes. Each proportion comes from one account and some hundred conversations, weighted in the direction of software program and AI instruments as a result of that’s what I ask about. A special question combine strikes the numbers.

My first cross was totally “finest X” queries, which meant I’d have revealed a declare about ChatGPT whereas solely testing one phrasing of 1 sort of query. The 24-query boundary check exists as a result of that bothered me, and it modified the discovering somewhat than confirming it.

Personalization is actual, and it confirmed up in my information. My meal package question got here again with UAE, Dubai, and an area firm. Journey insurance coverage went straight to a UAE insurer’s web site. I by no means stated the place I reside. So these are my account’s shortlists, not ChatGPT’s.

FanoutFox has a personalization label, so you may see in case your question is affected by personalization.

The 68.9% in opposition to 2.1% comparability makes use of complete conversations, so a model realized in flip one and queried in flip two counts as pre-known. That’s why the first-query check is the headline declare as a substitute. A 33 occasions hole is simply too huge for that contamination to clarify, however the clear quantity is the 21 of 27.

My model matching splits “Automobile and Driver” into two tokens and as soon as let “SaaS” by means of as a model. On the native question, it counted “Arabian Ranches” and “Dubai” as manufacturers, which they clearly aren’t. The queries themselves are the proof. Deal with my counts as approximate and skim the question strings as a substitute.

Examine It Your self In 2 Minutes

You don’t need to take my phrase for any of this, which is the entire level of studying visitors as a substitute of guessing.

1. Open ChatGPT in Chrome, open DevTools, go to the Community tab.
2. Ask “finest [your category] 2026”.
3. Filter for dialog, open the response, seek for queries.
4. Learn the primary question the mannequin wrote and search for names you by no means typed.

In case your rivals are in that checklist and also you’re not, now you recognize, and it took much less time than making a espresso.

That’s the guide model, and it’s high quality for one test. FanoutFox runs the identical test on each reply you give it, retains a historical past so you may watch a class shift over weeks, and by no means sends your conversations anyplace. It’s free, and it’s the instrument I exploit for all of this myself.

What’s Subsequent

The hole in all that is that it’s one account. I’ve requested my full ChatGPT information export to run the identical evaluation throughout each dialog I’ve ever had. The follow-up that will settle it’s the identical class sweep on a number of accounts without delay. If the identical manufacturers get injected for a similar questions on different individuals’s accounts, it’s model-level data. In the event that they don’t, personalization is doing extra work than anybody thinks.

Both reply is value having, and I’ll publish whichever one turns up. Something that strikes goes in the ChatGPT research tracker.

Learn off a logged-in ChatGPT Plus account in Dubai between 24 and 25 July 2026. 57 conversations for the quotation numbers, 27 for the first-query check, twelve contemporary class queries, and three repeats. Structural findings are stable at this pattern. Each proportion is directional.

Extra Assets:

This submit was initially revealed on Suganthan.

Featured Picture: Prostock-studio/Shutterstock