CITED.
THE GEO & AI CITATION LEDGER — DAILY
SEO got you ranked. GEO gets you cited.
This is not a blog you scroll. It is an answer engine you get quoted by.
SEO got you ranked. GEO gets you cited.
This is not a blog you scroll. It is an answer engine you get quoted by.
The click is dying as the unit of discovery. The manifesto for what replaces it, and why the brands that understand citation before 2027 will own their categories.
002Attention used to be measured in visits. Soon it is measured in how often the machines repeat your name. The new currency, measured across 21 live products.
003Ask the model about your company and read it as a stranger would. That answer is your real homepage now. The story of a citation tool that could not name itself, and the loop that fixes it.
004You could rent a ranking with content and links. You cannot rent a model's trust. Two real stories: a tool deindexed overnight that no technical fix could restore, and four flawless pages Google crawled and refused to index.
005Reputation engineering for machines that read everything and quote almost nothing. What GEO is not, the three levers that move citations, and the measured data on where AI citations actually come from.
006The anatomy of a citation magnet, with the real Bing numbers behind it: describability, the free paradox, and why machines recommend tools they can fully describe.
007What belongs in it, and the audit that caught mine lying. Why llms.txt earns near-zero citations directly but a wrong one gets you misquoted. Includes a copy-paste starter you can adapt.
008A free report almost nobody opens showed 1,500 Copilot citations I did not know I had. Where it hides, why only Bing has it, and the five step playbook.
009ChatGPT reads one of my sites 13 times a day and analytics never noticed. The six AI visitors, what each visit means, and the two commands that reveal them.
010Blocking AI crawlers in 2026 is deindexing yourself from the future. The VIP allow-list, why naming bots beats wildcard silence, and a copy-paste starter.
011Machines carry one sentence into an answer, never a paragraph. Three real rewrites from my own sites, and the five rules I write every line by.
012Position, answer box, or the model's memory. What each layer actually buys, with real citation data, and the order to fix them in.
013Zero impressions overnight with a perfectly clean dashboard. Two months of resubmitting failed. The 500 character ticket that restored it in two days, and the citations that came back 2.4 times stronger.
014The five properties that carry the weight, the two that hurt you, and the @id trick that turns scattered pages into one entity a model can hold.
015Models cross-check every source about you, and contradictions read as noise. Three contradictions I shipped on my own properties, and the six point cleanup checklist.
016I audited a site with a healthy traffic chart and ninety percent of it was machines. Seven days of a real server log, split three ways, and the five tests that separate scanners from answer engines from people.
017People search their fears. A free checker I nearly did not ship grew into 36 percent of a product’s search impressions, and the five rules for building the tool everyone is scared they need.
018Answers absorb the research visit and leave the buyer behind. Six numbers move when the tourists leave, and only one of them means what it says.
019Most AI crawlers never run JavaScript, and the live ones have seconds. The two number test that shows what your page ships before anything executes.
020Mobile LCP said 8.7 seconds on a page served in milliseconds. The headline was not slow, it was invisible, and the fix was one property.
021A model does not gather candidates and pick a winner. It recalls a reputation. Four observed patterns, each with a test you can run on your own brand.
022One property had pinged IndexNow exactly once, by hand, months earlier. The deploy hook, the file watcher, and why HTTP 202 is a success.
023Thirty one pages flagged, and every one was the site tidying address variants onto the canonical. Seven Search Console reports triaged into smoke, check and fire.
024Fifteen questions, three engines, once a month, answers recorded verbatim. The audit protocol for what AI tells strangers about your brand, and the three buckets every answer lands in.
025Five tools free at the core, every one of them named by machines, and the paid twins invisible. The access ladder from an open page to a hard paywall, and where the money still sits.
026A robots.txt fixed at the origin in one minute and served broken from the edge for two hours. The five copies of you, how long each one can stay wrong, and the habit that catches it.
027Adjectives cannot be checked, so they cannot be trusted, so they never travel. The floor rule, five worked examples of a soft claim hardened into a number, and where to put a figure so a machine finds it.
028Analytics runs in a browser and machines do not run it, so the readership that matters is missing from the chart by construction. One ordinary day drawn as a strip: three human hours, thirteen machine ones.
029A single domain can only ever be one witness testifying about itself. The vouch graph drawn two ways, and the four ties that have to match before separate domains resolve as one operation.
030Zero clicks, zero Bing impressions, 1,358 machine reads in fifteen days. The honest recap of thirty daily entries, including the three things in the crawl data I did not predict.
031robots.txt answers yes or no. llms.txt explains you. RSL states the terms, and it is the first file that can say yes to search and no to training separately.
0321,712 AI citations against 111 clicks, and almost none of them for the thing the site actually does. Read your citation queries, not just your totals.
033Three protocols now let software finish the purchase, and not one of them renders your page. The fields an agent actually reads, and the tier where half of it turns out to be policy.
034HTTP 402 came back from the dead so crawlers could be charged, and new domains now start closed by default. The three lanes a crawler meets at your edge, and which one you are in.
035An ad buys the slot, a citation buys the sentence, and only one survives the budget ending. Plus the fourth OpenAI crawler that almost nobody has put in their robots.txt.
036Four crawlers, four jobs, and two that never read your robots.txt. Blocking GPTBot costs you training, blocking OAI-SearchBot costs you today, and most sites block the wrong one.
037A citation is a mention, a tool call is a job. What the Model Context Protocol actually is, why the governance line decides whether to build on it, and the smallest version you can ship.
038Gartner says a quarter of searches this year and more than half by 2028. The forecast everybody quotes and nobody acts on, read from a site with no budget and no team.
039Almost half of AI citations land on listicles. Almost none of them are yours. How to get onto a list somebody else owns.
040One engine mentions brands twenty times more than another, and the difference is mostly Reddit. What that means if you have never posted there.
041YouTube is the most cited domain on ChatGPT and Perplexity. The reason has almost nothing to do with video.
042Wikipedia and Wikidata shape what every model believes about your category, and writing about yourself there is the fastest way to lose. The two surfaces beside it are open and almost nobody uses them.
043The cheapest untapped citation surface is the language you are not publishing in, and the numbers are not close.
044A review page about you carries roughly three times the citation weight of a page you wrote about yourself.
045Lists, threads, the encyclopaedia, reviews and transcripts write most AI citations. Line them up and one pattern survives: somebody else holds the pen.
046The engine with the most users learns you long before it quotes you. The training crawler arrived first, four to one. What to publish for memory.
047Fresh pages earn around 6 citations against 3.6 for stale ones, and its crawler went quiet the fortnight I stopped publishing. What freshness really means here.
048Claude searches through Brave, runs three crawlers that all obey robots.txt, and qualifies what it cannot support. How to become a source it will name.
049Google's two AI surfaces reach the same conclusion nine times in ten and share only 13.7% of their citations. Why they split, and what to write.
050Copilot grounds its answers in Bing, and Bing now reports every citation. How one small site went from 2 Copilot citations a day to 30, and how to repeat it.
051The same category question to five assistants returns five shortlists, because each searches a different pool. A twenty minute test, and how to read the gaps.
052ChatGPT sent a visit for every 2.6 fetches, Perplexity for about sixty, Claude for about 340. Readers and couriers, from my own server logs.
053Every crawler got full HTML, every check passed, and the site still sat between positions 20 and 60. Authority is the row no crawler fixes.
054Ask an assistant what happened to your company. If it answers in the past tense, it is reading old evidence. The tense check and the correction path.
05583 posts, 658 impressions, zero clicks, then two accounts suspended. Social was never distribution, and every citation I can trace landed on a page.
056Pages indexed, no errors in Webmaster Tools, and Bing impressions at zero. The 500 character support ticket that got a suppressed domain restored in two days.
057Research puts 44.2% of AI citations in the first 30% of a page. Your conclusion is in the wrong place, and moving it takes an afternoon.
058The page that names your competitors is the page that gets pulled into comparison answers. How to write alternatives pages that are fair, dated and quotable.
059A templated page set I planned to delete earned 46% of the clicks from 13% of the impressions. Outcome data beat the audit checklist.
060Twelve pages had full markup and the pricing page had none. Schema coverage is a list of the pages that matter, not a percentage.
061Three invisible bytes glued to two API keys took both AI providers down at once, and one error message blamed the request body.
062My caption tool invented words over silence, and my first guard deleted 25 real words. The threshold had to be measured on real speech.
063Assistants answer cost questions with whatever figure they can find. Publish the real number, mark it up with the right type and keep every copy in agreement.
064Not the newest and not the most official. Assistants tend to repeat the version that appears in more places, so correction is a counting job.
065A correction reaches a fraction of the people the mistake did. Four real mistakes, two that got out and two caught the night before.
066Lists collect 44% of AI citations, yet a small site wins by writing the explainer those lists depend on.
067Five predictions about AI search for 2027, each with a date and the outcome that would prove it wrong. Two I planned had already come true.
068Two months of real numbers measured like month one: total crawl held steady while live fetches from AI answers fell to zero.
069Nine days of logs from two MCP servers: 14,809 connections, 1,447 tool calls, and who actually wanted something done.
070What OpenAI asks for before a tool can appear inside ChatGPT: the server, the listing, eight test cases and the rules on selling, from a real submission.
071Crawled hundreds of times in fifteen days and fetched for a live answer zero times. The three stages of being read.
072Gate the download, never the answer. How to put a paywall on a page without hiding it from search and AI crawlers, with the markup that keeps it legitimate.
073Sitemaps update themselves; llms.txt does not. On one site it was eight pages behind, one a live article.
074Read only access to Search Console and Bing Webmaster data: the setup, what each API returns, how far behind each runs and how to keep the keys safe.
075Six months of Bing data from 41 sites: questions asking for a computed answer were clicked five times as often as explanations.
076A thirty word search asking for a JSON array, and one phrase in 24 variants. What machine written searches reveal.
077Search Console from launch to 9 October: 75 impressions, 2 clicks, and 37 of 63 pages never shown to anyone.
078What each review snippet warning in Search Console means, how to fix the nested itemReviewed conflict in minutes, and why self serving reviews never get stars.
079A link in our best guide in exchange for a feature with no date, page or topic. Three tests it failed.
08036 page loads said Google sent them; Google counted 4 clicks. Agent names that never fetched a page.
081Four posts credited our agency with a 23% and a 41% figure. Neither is on the page they link to.
082The same unchanged page scored 66, 83 and 99 in one hour. Medians, cached runs and missing field data.
083129 people used a product through ChatGPT; 14 ever opened its website and none saw the paywall.
084How Gemini grounds answers in the Google Search index, what the Google-Extended token really controls, which Google fetchers act for Gemini users and the steps that make a page citable.
085ChatGPT fetched one article 40 times in eleven days. Google ranks it 19th of 90 pages.
086CDNs skip HTML by default. One rule to cache pages safely, a purge on every deploy, browsers that revalidate, and what caching does to your crawler logs.
087Edge caching cut ChatGPT fetches in my server log by three quarters. The crawlers never left.
088ChatGPT read a careers page 42 times in two weeks and sent most of its visitors there to apply for closed roles.