{"id":23284,"date":"2026-08-14T22:51:01","date_gmt":"2026-08-14T22:51:01","guid":{"rendered":"https:\/\/scannn.com\/chatgpt-already-knows-who-itll-recommend-before-it-searches-suganthan\/"},"modified":"2026-08-14T22:51:01","modified_gmt":"2026-08-14T22:51:01","slug":"chatgpt-already-knows-who-itll-recommend-before-it-searches-suganthan","status":"publish","type":"post","link":"https:\/\/scannn.com\/lv\/chatgpt-already-knows-who-itll-recommend-before-it-searches-suganthan\/","title":{"rendered":"ChatGPT Already Knows Who It'll Recommend Before It Searches \u2014 Suganthan"},"content":{"rendered":"\n
\n

I asked ChatGPT for the best AI note taking app. Nothing in my question named a brand.<\/p>\n

Before it fetched anything, it wrote itself this search.<\/p>\n

best AI note taking apps 2026 official pricing features Granola Notion AI Otter Fireflies Fathom Mem Limitless<\/span><\/span><\/code><\/pre>\n
Read the tail of that string. Granola. Notion AI. Otter. Fireflies. Fathom. Mem. Limitless. Seven products in one search. I\u2019d named none of them<\/strong><\/em>, and nothing had come back from the web search yet, so no result put them there. Those names came out of the model.<\/p>\n
Then it ran nine more searches.<\/p>\n
site:granola.ai pricing features AI meeting notes 2026<\/span><\/span>\nsite:otter.ai pricing AI meeting notes 2026<\/span><\/span>\nsite:fathom.video pricing AI meeting assistant 2026<\/span><\/span>\nsite:notion.com product AI Meeting Notes official 2026 pricing<\/span><\/span>\nsite:fireflies.ai pricing official AI meeting notes 2026<\/span><\/span>\nsite:mem.ai pricing AI notes official 2026<\/span><\/span>\nsite:notebooklm.google official features pricing 2026<\/span><\/span><\/code><\/pre>\n
That\u2019s the fan-out I wrote about in the first two parts, and now you can see what it\u2019s made of. One search to name the shortlist, then one search per name on that shortlist, each pointed straight at that company\u2019s own website.<\/p>\n
So the fan-out was never a search for candidates, it was ChatGPT going down a list it already had, one name at a time.<\/p>\n
In the first two parts I wrote that you need to survive a site:yourdomain.com<\/code> probe, because ChatGPT runs them. It only runs one on you if you were named in that first search. Miss the shortlist and your website never gets looked at, however well built it is. The decision happens before anything touches your server.<\/p>\n
I\u2019ve spent two months reading this traffic for part 1<\/a> and part 2<\/a>. This is the first thing I\u2019ve found that changes what I\u2019d tell a client to do.<\/p>\n
Before you read any further, all of this comes off one account, so every percentage below is a direction rather than a measurement. The mechanism is a different matter. You can reproduce that on your own account in two minutes and I\u2019ll show you how at the end.<\/p>\n
The four ideas behind this<\/h2>\n
Each one sets up the next.<\/p>\n
    \n
  1. ChatGPT writes its own search queries, and you can read them.<\/li>\n
  2. Those queries already contain brands nobody mentioned.<\/li>\n
  3. Being in that query is worth about 33 times more than being findable.<\/li>\n
  4. Once you\u2019re in, a second and much harsher filter decides who gets cited.<\/li>\n<\/ol>\n
    Idea 1. It writes its own searches, and they\u2019re readable<\/h2>\n
    When you ask a question, ChatGPT rewrites it into search queries of its own, runs them, reads what comes back, then writes an answer.<\/p>\n
    Those queries sit in the response your browser downloads, under a key currently called search_queries<\/code>. OpenAI renamed it from search_model_queries<\/code> in early August 2026. Nothing about this is a leak. Your own browser needs that JSON to draw the page, and you can read it in DevTools on your own account in about two minutes. Recipe\u2019s at the end.<\/p>\n
    Here\u2019s one from a question about live chat software.<\/p>\n
    best AI live chat customer support software 2026<\/span><\/span>\nIntercom Fin Zendesk AI pricing official<\/span><\/span><\/code><\/pre>\n
    I asked for \u201cbest ai based live chat support software\u201d. It added the year, \u201cofficial\u201d, \u201cpricing\u201d and three names.<\/p>\n
    Everything below comes from reading a few hundred of these.<\/p>\n
    Start here.<\/strong> Open ChatGPT, ask the \u201cbest [your category]\u201d question your buyers ask, and read the query it writes back. That one string shows you whether ChatGPT knows your brand exists, which is what AI visibility audits are sold to answer. Everything else in this article is about what to do with what you find in it.<\/p>\n
    Idea 2. The shortlist exists before the search runs<\/h2>\n
    The obvious objection to my note taking example is timing. Maybe ChatGPT searched once, saw those brands, then wrote a smarter second query. That\u2019d make the names a result of retrieval rather than a cause.<\/p>\n
    So I tested it properly. For each conversation I took the first user message and the first search query, ordered by timestamp. At that point nothing\u2019s been fetched, so there\u2019s no earlier result to have learned from.<\/p>\n
    In 21 of 27 conversations, that first query contained brands the user never typed.<\/strong><\/p>\n
    Look at rows two and three. Same question, slightly different phrasing, and the list grew from three names to seven. The list stretches with how you ask instead of coming out of a fixed table.<\/p>\n
    Then I ran twelve categories that have nothing to do with each other, to check this wasn\u2019t a software thing.<\/p>\n
    11 of 13 did the same.<\/strong><\/p>\n
    The robot vacuum row is the one that got me. Recalling that Roborock exists would be unremarkable. It recalled Saros<\/strong>, Dreame X50<\/strong> and Eufy S1 Pro<\/strong>. Current model numbers, in the first query, unprompted. Whatever this knowledge is, it goes down to the product line.<\/p>\n
    The electric SUV row does something different from all 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 one category its instinct is to go to the makers, in another the reviewers.<\/p>\n
    I got excited about that and then re-ran three categories to see if any of it holds still.<\/p>\n
    Language learning barely moved. Five of six names came back the same. Accounting collapsed from six vendors to a single targeted probe at QuickBooks. Web hosting switched sides completely, dropped every vendor and went to a review site instead.<\/p>\n
    So vendors versus magazines is a tendency that can flip between runs, not a fixed property of the category. Two things survive that. The injection happened every single time, and the category with the clearest market leaders kept its names. My guess is that settled categories have stable shortlists and contested ones wobble, but three repeats isn\u2019t enough to say that.<\/p>\n
    Don\u2019t judge your AI visibility from one answer. Run the question five times, because the list changes between runs.<\/p>\n
    It happens whenever ChatGPT has to supply the products<\/h3>\n
    Every query up to this point had the word \u201cbest\u201d in it, so that was the first thing I tried to break. I ran twenty-four more queries that avoided the word, across seven different shapes.<\/p>\n
    \u201cBest\u201d turned out to have nothing to do with it. What matters is whether ChatGPT has to come up with the products itself.<\/p>\n
    When it doesn\u2019t search at all, none of this applies.<\/strong> Ask how noise cancelling works, or what a vector database is, and it answers from training with no web search. Same for open-ended moaning. \u201cWe are spending too much on customer support tooling\u201d got no search either. Seven of my twenty-four never touched the web, so there was no shortlist to get onto.<\/p>\n
    Name the brands yourself and it takes them.<\/strong> \u201cXero or QuickBooks for a small business\u201d went straight to site:xero.com\/uk<\/code>. \u201cShould I use HubSpot for a small agency\u201d went to site:hubspot.com<\/code>.<\/p>\n
    Leave the candidates to it and it reaches into memory.<\/strong> That happened in ten of the eleven cases where I asked for a recommendation without naming anything.<\/p>\n
    The third row is my favourite. I deliberately avoided saying \u201crobot vacuum\u201d and named nobody. It went to a specific Roborock model\u2019s own website on the first search.<\/p>\n
    Row two should worry you if you sell software. That\u2019s a complaint, not a shopping question. Somebody grumbling about their meeting notes got turned into Granola official pricing<\/code> before a single page was fetched.<\/p>\n
    And there\u2019s a version of this aimed straight at your competitors. All three of my displacement queries pulled in new names. \u201cAlternatives to zendesk\u201d produced Help Scout, \u201cwhat can i use instead of quickbooks\u201d produced Zoho Books, and \u201csomething like duolingo but better for grammar\u201d produced Kwiziq and Babbel. When your customer goes looking for a way out of your rival, ChatGPT nominates whoever it already knows. That could be you, and you don\u2019t get a say in it on the day.<\/p>\n
    If ChatGPT searches for a product and you haven\u2019t named one, it brings its own.<\/strong><\/p>\n
    Do this.<\/strong> Run your category question five times and write down the names that appear in the query each time. The names that show up every run are your real competitive set in ChatGPT\u2019s head. Anything that comes and goes is contested ground, which is where a push can move something. If your brand never appears in five runs, you have your answer and it isn\u2019t a technical one.<\/p>\n
    Idea 3. Being in the query is worth about 33 times more<\/h2>\n
    I sorted every brand into two groups. Ones that appeared in a query ChatGPT wrote, and ones that were only fetched during the search without ever being named. Then I checked how often each group made it into the final answer.<\/p>\n
    About 33 times the difference.<\/p>\n
    I also found 86 cases where a brand got recommended and its website was never fetched in that conversation at all. A mention doesn\u2019t need a crawl.<\/p>\n
    This is the uncomfortable part for my own industry. Most of what\u2019s sold as GEO right now is retrieval work, which is the 2.1% column. The 68.9% column is decided before any of it runs.<\/p>\n
    Do this.<\/strong> Split your budget to match the two columns. If you\u2019re absent from the query, the money belongs in the work that gets you written about, reviewed, compared and listed. Digital PR, category content, review-site placement, analyst coverage, showing up in the roundups your buyers read. If you\u2019re already in the query, that spend is largely done and the technical work below is what\u2019s left.<\/p>\n
    Idea 4. Being in the query is the entry ticket, not the win<\/h2>\n
    If it stopped there the advice would be \u201cbuild brand equity\u201d and we could all go home. A second filter runs after the query, and it\u2019s brutal.<\/p>\n
    I built a labelled dataset from 57 conversations. Every retrieved page as a row, with whether it earned a citation as the label. 3,554 pages.<\/p>\n
    110 got cited. That\u2019s 3.1%.<\/strong><\/p>\n
    ChatGPT reads around 600 pages to write one answer and credits about 30. Almost everything gets read. The gap between being read and being credited is where the work is.<\/p>\n
    Three things separated the cited from the ignored.<\/p>\n
    Position, hard.<\/strong> ChatGPT groups results by domain, and where you sit in that group predicts nearly everything.<\/p>\n
    Below the top two, citation is a rounding error. Being in the retrieved set isn\u2019t a win if you\u2019re ninth.<\/p>\n
    Piling on pages hurts.<\/strong> When several pages from one domain show up in the same group, per-page conversion collapses.<\/p>\n
    Two tightly matched pages is the sweet spot. Past six you\u2019re mostly competing with yourself. I\u2019ve been telling clients to consolidate for a year on instinct. This is the first time I\u2019ve watched it happen in the data.<\/p>\n
    Relevance qualifies you, it doesn\u2019t select you.<\/strong> I scored the cited page against every other page retrieved for the same claim, on how well its text matched the sentence being supported. The cited page sat in the top 5% of the pool. But it was the single best match only 20% of the time, and its average overlap was well below the best available.<\/p>\n
    So claim relevance builds a shortlist and something else picks the winner. Get into the top 10% for a specific claim and you\u2019re in the conversation. That part you control by writing. The final pick involves things I can\u2019t see.<\/p>\n
    Do this.<\/strong> Take the questions your buyers ask and find every page of yours that answers the same one. That\u2019s the bottom row of the table above, and it\u2019s costing you. Pick the page that matches the intent most tightly, make it the answer, and fold or redirect the rest into it. Then make sure the sentence that actually answers the question sits near the top of that page in plain HTML text, with the numbers in it.<\/p>\n
    Finding those overlaps by hand across a real site is the miserable part. Cannibalisation is normally judged on keyword overlap, and that\u2019s the wrong unit here, because ChatGPT groups by claim rather than by keyword.<\/p>\n
    This is what Keyword Insights<\/a> was built for.<\/p>\n
    It clusters your keywords by search intent rather than by string match, so the pages fighting over one intent surface as a group even when they share no keywords.<\/p>\n
    That grouping maps closely onto what I\u2019m watching ChatGPT do when it collapses a domain down to one cited page.<\/p>\n
    <\/p>\n
    Full disclosure, it\u2019s my company, which is also why I know it handles this specific job.<\/p>\n
    What I\u2019d actually do with this<\/h2>\n
    The findings split AI visibility into two games that keep getting treated as one.<\/p>\n
    Game one is being in the category vocabulary<\/h3>\n
    If ChatGPT doesn\u2019t already connect your brand to your category, it won\u2019t name you in the query, and you\u2019re looking at a 2% chance of a mention.<\/p>\n
    Schema won\u2019t fix that, and neither will page speed. An llms.txt<\/code> file has even less of a chance, because your server never gets contacted before the decision\u2019s made.<\/p>\n
    What seems to build it is slow and unglamorous. Being written about, reviewed, compared and argued over across the open web for years, until the association exists in the training data.<\/p>\n
    That\u2019s digital PR and category-defining content, which is awkward for everyone currently selling technical audits as an AI strategy.<\/p>\n
    Game two is winning the citation once you\u2019re in<\/h3>\n
    That\u2019s the 3.1%, and it\u2019s real. One tightly matched page per intent, the claim-bearing sentence early, facts and numbers in plain HTML text, and no cluster of near-identical pages fighting each other.<\/p>\n
    Here\u2019s the whole thing as a sequence you can run this week.<\/p>\n
      \n
    1. \n
      Read the query.<\/strong> Ask your category question five times and note the brands ChatGPT writes into its own search. Ten minutes.<\/p>\n<\/li>\n
    2. \n
      Find which game you\u2019re in.<\/strong> In the query every time means you\u2019re playing game two. Never in it means game one, and no amount of page work will move you.<\/p>\n<\/li>\n
    3. \n
      If you\u2019re absent, go and get written about.<\/strong> Reviews, roundups, comparison pages, the publications your category reads. In a category like electric SUVs, where ChatGPT went to the magazines before it went to any manufacturer, that\u2019s the only door.<\/p>\n<\/li>\n
    4. \n
      If you\u2019re present, check conversion not retrieval.<\/strong> Count how often your pages get fetched against how often they get credited. Fetched a lot and credited rarely is a page problem, not a visibility problem.<\/p>\n<\/li>\n
    5. \n
      Fix the page, not the site.<\/strong> One page per intent, the answer sentence near the top, real numbers in HTML text rather than in an image or loaded by JavaScript.<\/p>\n<\/li>\n
    6. \n
      Re-check monthly.<\/strong> The stability test showed the shortlist moves, and the format moves with it. In early August 2026 OpenAI renamed the key these queries sit under, and the fan-outs in my captures dropped from 12 searches per answer to 4. What was true in July won\u2019t necessarily hold in September.<\/p>\n<\/li>\n<\/ol>\n
      In my data one brand got fetched 66 times and never cited once. That\u2019s not an awareness problem. The engine kept going back and kept deciding they had nothing worth quoting. Step four exists to catch that.<\/p>\n
      Steps one and four are what I built FanoutFox<\/a> for, because running them by hand for every answer gets old fast. It\u2019s a free Chrome extension, everything stays in your browser, and it reads your own ChatGPT session rather than guessing from the outside. It shows you the fan-out queries verbatim, so step one is a glance instead of a DevTools session.<\/p>\n
      \"\"<\/p>\n
      For step four it now has a converts column, retrieved against cited per domain, the number this article argues you should be tracking instead of retrieval counts. A domain that gets read repeatedly and never credited is the diagnosis you\u2019re looking for, and it\u2019s sorted right there.<\/p>\n
      The caveats<\/h2>\n
      I\u2019d rather you trust the parts that deserve it.<\/p>\n
      The mechanism is solid. ChatGPT writing brand names into its first query is visible in a single capture and you can reproduce it yourself in two minutes. Every percentage comes from one account and a few hundred conversations, weighted towards software and AI tools because that\u2019s what I ask about. A different query mix moves the numbers.<\/p>\n
      My first pass was entirely \u201cbest X\u201d queries, which meant I\u2019d have published a claim about ChatGPT while only testing one phrasing of one kind of question. The twenty-four query boundary test exists because that bothered me, and it changed the finding rather than confirming it.<\/p>\n
      Personalisation is real and it showed up in my data. My meal kit query came back with UAE<\/code>, Dubai<\/code> and a local company. Travel insurance went straight to a UAE insurer\u2019s site. I never said where I live. So these are my account\u2019s shortlists, not ChatGPT\u2019s.<\/p>\n
      FanoutFox has a personalisation label, so you can see if your query is affected by personalisation.<\/p>\n
      \"\"<\/p>\n
      The 68.9% against 2.1% comparison uses whole conversations, so a brand learned in turn one and queried in turn two counts as pre-known. That\u2019s why the first-query test is the headline claim instead. A 33 times gap is too big for that contamination to explain, but the clean number is the 21 of 27.<\/p>\n
      My brand matching splits \u201cCar and Driver\u201d into two tokens and once let \u201cSaaS\u201d through as a brand. On the local query it counted \u201cArabian Ranches\u201d and \u201cDubai\u201d as brands, which they obviously aren\u2019t. The queries themselves are the evidence. Treat my counts as approximate and read the query strings instead.<\/p>\n
      Check it yourself in two minutes<\/h2>\n
      You don\u2019t have to take my word for any of this, which is the whole point of reading traffic instead of guessing.<\/p>\n
        \n
      1. Open ChatGPT in Chrome, open DevTools, go to the Network tab.<\/li>\n
      2. Ask \u201cbest [your category] 2026\u201d.<\/li>\n
      3. Filter for conversation<\/code>, open the response, search for queries<\/code>.<\/li>\n
      4. Read the first query the model wrote and look for names you never typed.<\/li>\n<\/ol>\n
        If your competitors are in that list and you\u2019re not, now you know, and it took less time than making a coffee.<\/p>\n
        That\u2019s the manual version, and it\u2019s fine for one check. FanoutFox<\/a> runs the same check on every answer you give it, keeps a history so you can watch a category shift over weeks, and never sends your conversations anywhere. It\u2019s free and it\u2019s the tool I use for all of this myself.<\/p>\n
        \"\"<\/p>\n
        What\u2019s next<\/h2>\n
        The gap in all this is that it\u2019s one account. I\u2019ve requested my full ChatGPT data export to run the same analysis across every conversation I\u2019ve ever had. The follow-up that would settle it is the same category sweep on several accounts at once. If the same brands get injected for the same questions on other people\u2019s accounts, it\u2019s model-level knowledge. If they don\u2019t, personalisation is doing more work than anyone thinks.<\/p>\n
        Either answer is worth having, and I\u2019ll publish whichever one turns up. Anything that moves goes in the ChatGPT research tracker<\/a>.<\/p>\n
        Read off a logged-in ChatGPT Plus account in Dubai between 24 and 25 July 2026. 57 conversations for the citation numbers, 27 for the first-query test, twelve fresh category queries and three repeats. Structural findings are solid at this sample. Every percentage is directional.<\/em><\/p>\n<\/p><\/div>\n
        Source link <\/a><\/p>\n","protected":false},"excerpt":{"rendered":"
        I asked ChatGPT for the best AI note taking app. Nothing in my question named a brand. Before it fetched anything, it wrote itself this search. best AI note taking apps 2026 official pricing features Granola Notion AI Otter Fireflies Fathom Mem Limitless Read the tail of that string. Granola. Notion AI. Otter. Fireflies. Fathom. […]<\/p>\n","protected":false},"author":16,"featured_media":23285,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[143],"tags":[],"class_list":["post-23284","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ai"],"_links":{"self":[{"href":"https:\/\/scannn.com\/lv\/wp-json\/wp\/v2\/posts\/23284","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/scannn.com\/lv\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/scannn.com\/lv\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/scannn.com\/lv\/wp-json\/wp\/v2\/users\/16"}],"replies":[{"embeddable":true,"href":"https:\/\/scannn.com\/lv\/wp-json\/wp\/v2\/comments?post=23284"}],"version-history":[{"count":0,"href":"https:\/\/scannn.com\/lv\/wp-json\/wp\/v2\/posts\/23284\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/scannn.com\/lv\/wp-json\/wp\/v2\/media\/23285"}],"wp:attachment":[{"href":"https:\/\/scannn.com\/lv\/wp-json\/wp\/v2\/media?parent=23284"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/scannn.com\/lv\/wp-json\/wp\/v2\/categories?post=23284"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/scannn.com\/lv\/wp-json\/wp\/v2\/tags?post=23284"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}