ChatGPT Already Knows Who It’ll Recommend Before It Searches

I asked ChatGPT for the champion AI note-taking app. Seven words, nary marque names.

Before it fetched anything, it wrote itself this search.

best AI statement taking apps 2026 charismatic pricing features Granola Notion AI Otter Fireflies Fathom Mem Limitless

Read the tail of that string. Granola. Notion AI. Otter. Fireflies. Fathom. Mem. Limitless. Seven products successful 1 search. I’d named nary of them, and thing had travel backmost from the web hunt yet, truthful nary consequence put them there. Those names came retired of the model.

Then it ran 9 much searches.

site:granola.ai pricing features AI gathering notes 2026 site:otter.ai pricing AI gathering notes 2026 site:fathom.video pricing AI gathering adjunct 2026 site:notion.com merchandise AI Meeting Notes charismatic 2026 pricing site:fireflies.ai pricing charismatic AI gathering notes 2026 site:mem.ai pricing AI notes charismatic 2026 site:notebooklm.google charismatic features pricing 2026

That’s the fan-out I wrote astir successful the first 2 parts, and now you tin spot what it’s made of. One hunt to sanction the shortlist, past 1 hunt per sanction connected that shortlist, each pointed consecutive astatine that company’s ain website.

So the fan-out was ne'er a hunt for candidates; it was ChatGPT going down a database it already had, 1 sanction astatine a time.

In the first 2 parts, I wrote that you request to past a site:yourdomain.com probe, because ChatGPT runs them. It only runs 1 connected you if you were named successful that first search. Miss the shortlist and your website ne'er gets looked at, nevertheless good built it is. The determination happens earlier thing touches your server.

I’ve spent 2 months reference this postulation for part 1 and part 2. This is the first point I’ve recovered that changes what I’d show a customer to do.

Before you publication immoderate further, each of this comes disconnected 1 account, truthful each percent beneath is simply a guidance alternatively than a measurement. The system is simply a different matter. You tin reproduce that connected your ain relationship successful 2 minutes, and I’ll show you really astatine the end.

The 4 Ideas Behind This

Each 1 sets up the next.

1. ChatGPT writes its ain hunt queries, and you tin publication them.
2. Those queries already incorporate brands cipher mentioned.
3. Being successful that query is worthy astir 33 times much than being findable.
4. Once you’re in, a 2nd and overmuch harsher select decides who gets cited.

Idea 1: It Writes Its Own Searches, And They’re Readable

When you inquire a question, ChatGPT rewrites it into search queries of its own, runs them, sounds what comes back, past writes an answer.

Those queries beryllium successful the consequence your browser downloads, nether a cardinal presently called search_queries. OpenAI renamed it from search_model_queries successful early August 2026. Nothing astir this is simply a leak. Your ain browser needs that JSON to tie the page, and you tin publication it successful DevTools connected your ain relationship successful astir 2 minutes. Recipe’s astatine the end.

Here’s 1 from a mobility astir unrecorded chat software.

best AI unrecorded chat customer support package 2026 Intercom Fin Zendesk AI pricing official

I asked for “best ai based unrecorded chat support software.” It added the year, “official,” “pricing,” and 3 names.

Everything beneath comes from reference a fewer 100 of these.

Start here. Open ChatGPT, inquire the “best [your category]” mobility your buyers ask, and publication the query it writes back. That 1 drawstring shows you whether ChatGPT knows your marque exists, which is what AI visibility audits are sold to answer. Everything other successful this article is astir what to do pinch what you find successful it.

Idea 2: The Shortlist Exists Before The Search Runs

The evident objection to my note-taking illustration is timing. Maybe ChatGPT searched once, saw those brands, past wrote a smarter 2nd query. That’d make the names a consequence of retrieval alternatively than a cause.

So I tested it properly. For each conversation, I took the first personification connection and the first hunt query, ordered by timestamp. At that constituent nothing’s been fetched, truthful there’s nary earlier consequence to person learned from.

In 21 of 27 conversations, that first query contained brands the personification ne'er typed.

Look astatine rows 2 and three. Same question, somewhat different phrasing, and the database grew from 3 names to seven. The database stretches pinch really you inquire alternatively of coming retired of a fixed table.

Then I ran  12 categories that person thing to do pinch each other, to cheque this wasn’t a package thing.

11 of 13 did the same.

The robot vacuum statement is the 1 that sewage me. Recalling that Roborock exists would beryllium unremarkable. It recalled Saros, Dreame X50, and Eufy S1 Pro. Current exemplary numbers, successful the first query, unprompted. Whatever this knowledge is, it goes down to the merchandise line.

The electrical SUV statement does thing different from each the others. For accounting and therapy and hosting, it named vendors and went to their pricing pages. For cars, it named magazines. Car and Driver, Edmunds, Top Gear. In 1 category, its small heart is to spell to the makers; successful another, the reviewers.

I sewage excited astir that and past re-ran 3 categories to spot if immoderate of it holds still.

Language learning hardly moved. Five of six names came backmost the same. Accounting collapsed from six vendors to a azygous targeted probe astatine QuickBooks. Web hosting switched sides completely, dropped each vendor and went to a reappraisal tract instead.

So vendors versus magazines is simply a inclination that tin flip betwixt runs, not a fixed spot of the category. Two things past that. The injection happened each azygous time, and the class pinch the clearest marketplace leaders kept its names. My conjecture is that settled categories person unchangeable shortlists and contested ones wobble, but 3 repeats isn’t capable to opportunity that.

Don’t judge your AI visibility from 1 answer. Run the mobility 5 times, because the database changes betwixt runs.

It Happens Whenever ChatGPT Has To Supply The Products

Every query up to this constituent had the connection “best” successful it, truthful that was the first point I tried to break. I ran twenty-four much queries that avoided the word, crossed 7 different shapes.

“Best” turned retired to person thing to do pinch it. What matters is whether ChatGPT has to travel up pinch the products itself.

When it doesn’t hunt astatine all, nary of this applies. Ask really noise-cancelling works, aliases what a vector database is, and it answers from training pinch nary web search. Same for open-ended moaning. “We are spending excessively overmuch connected customer support tooling” sewage nary hunt either. Seven of my twenty-four ne'er touched the web, truthful location was nary shortlist to get onto.

Name the brands yourself, and it takes them. “Xero aliases QuickBooks for a mini business” went consecutive to site:xero.com/uk. “Should I usage HubSpot for a mini agency” went to site:hubspot.com.

Leave the candidates to it, and it reaches into memory. That happened successful 10 of the eleven cases wherever I asked for a proposal without naming anything.

The 3rd statement is my favorite. I deliberately avoided saying “robot vacuum” and named nobody. It went to a circumstantial Roborock model’s ain website connected the first search.

Row 2 should interest you if you waste software. That’s a complaint, not a shopping question. Somebody grumbling astir their gathering notes sewage turned into Granola charismatic pricing earlier a azygous page was fetched.

And there’s a type of this aimed consecutive astatine your competitors. All 3 of my displacement queries pulled successful caller names. “Alternatives to Zendesk” produced Help Scout; “what tin one usage alternatively of quickbooks” produced Zoho Books; and “something for illustration duolingo but amended for grammar” produced Kwiziq and Babbel. When your customer goes looking for a measurement retired of your rival, ChatGPT nominates whoever it already knows. That could beryllium you, and you don’t get a opportunity successful it connected the day.

If ChatGPT searches for a merchandise and you haven’t named one, it brings its own.

Do this. Run your class mobility 5 times and constitute down the names that look successful the query each time. The names that show up each tally are your existent competitory group successful ChatGPT’s head. Anything that comes and goes is contested ground, which is wherever a push tin move something. If your marque ne'er appears successful 5 runs, you person your answer, and it isn’t a method one.

Idea 3: Being In The Query Is Worth About 33 Times More

I sorted each marque into 2 groups. Ones that appeared successful a query ChatGPT wrote, and ones that were only fetched during the hunt without ever being named. Then I checked really often each group made it into the last answer.

About 33 times the difference.

I besides recovered 86 cases wherever a marque sewage recommended, and its website was ne'er fetched successful that speech astatine all. A mention doesn’t request a crawl.

This is the uncomfortable portion for my ain industry. Most of what’s sold arsenic GEO correct now is retrieval work, which is the 2.1% column. The 68.9% file is decided earlier immoderate of it runs.

Do this. Split your fund to lucifer the 2 columns. If you’re absent from the query, the money belongs successful the activity that gets you written about, reviewed, compared, and listed. Digital PR, class content, review-site placement, expert coverage, showing up successful the roundups your buyers read. If you’re already successful the query, that walk is mostly done, and the method activity beneath is what’s left.

Idea 4: Being In The Query Is The Entry Ticket, Not The Win

If it stopped there, the proposal would beryllium “build marque equity,” and we could each spell home. A 2nd select runs aft the query, and it’s brutal.

I built a labelled dataset from 57 conversations. Every retrieved page arsenic a row, pinch whether it earned a citation arsenic the label. 3,554 pages.

110 sewage cited. That’s 3.1%.

ChatGPT sounds astir 600 pages to constitute 1 reply and credits astir 30. Almost everything gets read. The spread betwixt being publication and being credited is wherever the activity is.

Three things separated the cited from the ignored.

Position, hard. ChatGPT groups results by domain, and wherever you beryllium successful that group predicts astir everything.

Below the apical two, citation is simply a rounding error. Being successful the retrieved group isn’t a triumph if you’re ninth.

Piling connected pages hurts. When respective pages from 1 domain show up successful the aforesaid group, per-page conversion collapses.

Two tightly matched pages is the saccharine spot. Past six, you’re mostly competing pinch yourself. I’ve been telling clients to consolidate for a twelvemonth connected instinct. This is the first clip I’ve watched it hap successful the data.

Relevance qualifies you; it doesn’t prime you. I scored the cited page against each different page retrieved for the aforesaid claim, connected really good its matter matched the condemnation being supported. The cited page sat successful the apical 5% of the pool. But it was the azygous champion lucifer only 20% of the time, and its mean overlap was good beneath the champion available.

So declare relevance builds a shortlist and thing other picks the winner. Get into the apical 10% for a circumstantial claim, and you’re successful the conversation. That portion you power by writing. The last prime involves things I can’t see.

Do this. Take the questions your buyers inquire and find each page of yours that answers the aforesaid one. That’s the bottommost statement of the array above, and it’s costing you. Pick the page that matches the intent astir tightly, make it the answer, and fold aliases redirect the remainder into it. Then make judge the condemnation that really answers the mobility sits adjacent the apical of that page successful plain HTML text, pinch the numbers successful it.

Finding those overlaps by manus crossed a existent tract is the miserable part. Cannibalization is usually judged connected keyword overlap, and that’s the incorrect portion here, because ChatGPT groups by declare alternatively than by keyword.

This is what Keyword Insights was built for.

It clusters your keywords by hunt intent alternatively than by drawstring match, truthful the pages fighting complete 1 intent aboveground arsenic a group moreover erstwhile they stock nary keywords.

That grouping maps intimately onto what I’m watching ChatGPT do erstwhile it collapses a domain down to 1 cited page.

Image Credit: Suganthan Mohanadasan

Full disclosure: It’s my company, which is besides why I cognize it handles this circumstantial job.

What I’d Actually Do With This

The findings divided AI visibility into 2 games that support getting treated arsenic one.

Game One Is Being In The Category Vocabulary

If ChatGPT doesn’t already link your marque to your category, it won’t sanction you successful the query, and you’re looking astatine a 2% chance of a mention.

Schema won’t hole that, and neither will page speed. An llms.txt file has moreover little of a chance, because your server ne'er gets contacted earlier the decision’s made.

What seems to build it is slow and unglamorous. Being written about, reviewed, compared, and based on complete crossed the unfastened web for years, until the relation exists successful the training data.

That’s integer PR and category-defining content, which is awkward for everyone presently trading method audits arsenic an AI strategy.

Game Two Is Winning The Citation Once You’re In

That’s the 3.1%, and it’s real. One tightly matched page per intent, the claim-bearing condemnation early, facts and numbers successful plain HTML text, and nary cluster of near-identical pages fighting each other.

Here’s the full point arsenic a series you tin tally this week.

1. Read the query. Ask your class mobility 5 times and statement the brands ChatGPT writes into its ain search. Ten minutes.
2. Find which crippled you’re in. In the query each clip intends you’re playing crippled two. Never successful it intends crippled one, and nary magnitude of page activity will move you.
3. If you’re absent, spell and get written about. Reviews, roundups, comparison pages, the publications your class reads. In a class for illustration electrical SUVs, wherever ChatGPT went to the magazines earlier it went to immoderate manufacturer, that’s the only door.
4. If you’re present, cheque conversion, not retrieval. Count really often your pages get fetched against really often they get credited. Fetched a batch and credited seldom is simply a page problem, not a visibility problem.
5. Fix the page, not the site. One page per intent, the reply condemnation adjacent the top, existent numbers successful HTML matter alternatively than successful an image aliases loaded by JavaScript.
6. Re-check monthly. The stableness trial showed the shortlist moves, and the format moves pinch it. In early August 2026, OpenAI renamed the cardinal these queries beryllium under, and the fan-outs successful my captures dropped from 12 searches per reply to 4. What was existent successful July won’t needfully clasp successful September.

In my data, 1 marque sewage fetched 66 times and ne'er cited once. That’s not an consciousness problem. The motor kept going backmost and kept deciding they had thing worthy quoting. Step 4 exists to drawback that.

Steps 1 and 4 are what I built FanoutFox for, because moving them by manus for each reply gets aged fast. It’s a free Chrome extension; everything stays successful your browser, and it sounds your ain ChatGPT convention alternatively than guessing from the outside. It shows you the fan-out queries verbatim, truthful measurement 1 is simply a glimpse alternatively of a DevTools session.

Image Credit: Suganthan Mohanadasan

For measurement four, it now has a converts column, retrieved against cited per domain, the number this article argues you should beryllium search alternatively of retrieval counts. A domain that gets publication many times and ne'er credited is the test you’re looking for, and it’s sorted correct there.

The Caveats

I’d alternatively you spot the parts that merit it.

The system is solid. ChatGPT penning marque names into its first query is visible successful a azygous capture, and you tin reproduce it yourself successful 2 minutes. Every percent comes from 1 relationship and a fewer 100 conversations, weighted towards package and AI devices because that’s what I inquire about. A different query operation moves the numbers.

My first walk was wholly “best X” queries, which meant I’d person published a declare astir ChatGPT while only testing 1 phrasing of 1 benignant of question. The 24-query bound trial exists because that bothered me, and it changed the uncovering alternatively than confirming it.

Personalization is real, and it showed up successful my data. My repast kit query came backmost pinch UAE, Dubai, and a section company. Travel security went consecutive to a UAE insurer’s site. I ne'er said wherever I live. So these are my account’s shortlists, not ChatGPT’s.

FanoutFox has a personalization label, truthful you tin spot if your query is affected by personalization.

Image Credit: Suganthan Mohanadasan

The 68.9% against 2.1% comparison uses full conversations, truthful a marque learned successful move 1 and queried successful move 2 counts arsenic pre-known. That’s why the first-query trial is the header declare instead. A 33 times spread is excessively large for that contamination to explain, but the cleanable number is the 21 of 27.

My marque matching splits “Car and Driver” into 2 tokens and erstwhile fto “SaaS” done arsenic a brand. On the section query, it counted “Arabian Ranches” and “Dubai” arsenic brands, which they evidently aren’t. The queries themselves are the evidence. Treat my counts arsenic approximate and publication the query strings instead.

Check It Yourself In 2 Minutes

You don’t person to return my connection for immoderate of this, which is the full constituent of reference postulation alternatively of guessing.

1. Open ChatGPT successful Chrome, unfastened DevTools, spell to the Network tab.
2. Ask “best [your category] 2026”.
3. Filter for conversation, unfastened the response, hunt for queries.
4. Read the first query the exemplary wrote and look for names you ne'er typed.

If your competitors are successful that database and you’re not, now you know, and it took little clip than making a coffee.

That’s the manual version, and it’s good for 1 check. FanoutFox runs the aforesaid cheque connected each reply you springiness it, keeps a history truthful you tin watch a class displacement complete weeks, and ne'er sends your conversations anywhere. It’s free, and it’s the instrumentality I usage for each of this myself.

Image Credit: Suganthan Mohanadasan

What’s Next

The spread successful each this is that it’s 1 account. I’ve requested my afloat ChatGPT information export to tally the aforesaid study crossed each speech I’ve ever had. The follow-up that would settee it is the aforesaid class expanse connected respective accounts astatine once. If the aforesaid brands get injected for the aforesaid questions connected different people’s accounts, it’s model-level knowledge. If they don’t, personalization is doing much activity than anyone thinks.

Either reply is worthy having, and I’ll people whichever 1 turns up. Anything that moves goes successful the ChatGPT investigation tracker.

Read disconnected a logged-in ChatGPT Plus relationship successful Dubai betwixt 24 and 25 July 2026. 57 conversations for the citation numbers, 27 for the first-query test, 12 caller class queries, and 3 repeats. Structural findings are coagulated astatine this sample. Every percent is directional.

More Resources:

• How ChatGPT Actually Picks Sources (I Read The Network Traffic, Not The Outputs)
• AI Visibility Rankings Aren’t Stable – New Research Shows It’s Mostly Statistical Noise
• Why Publishing More Content Is Making Your SEO Worse

---

This station was primitively published connected Suganthan.

---

Featured Image: Prostock-studio/Shutterstock