SEO has ‘died’ three times since I started doing it, I’m starting to lose count. Hummingbird killed it in September 2013. Then mobile killed it in April 2015. Then ChatGPT killed it in November 2022, and these days it dies again most Tuesdays on LinkedIn, usually in a post from a GEO agency that’s younger than the ChatGPT ads program. There’s always a merch table at the funeral too. This time around, it’s an AI visibility score, $250 a month on the entry plan, and apparently your brand got a 23.
Here’s the thing I want to be clear about before we go any further. LLM SEO is a real job, and I do it. That score, though? That’s a screenshot of one prompt, typed once, on one afternoon.
Back in November 2025, Ahrefs went and re-ran 43,000 AI Overview queries and found the answer rewrites itself about every two days, with close to half the cited sources swapped out every time it does. We’ve had the better part of a year to learn that. We’ve also had the better part of a year to build dashboards that hide it, and plenty of people chose the dashboard.
I’ve been doing SEO for 13 years now, and the site AI Overviews cite more than any other right now, YouTube, is the one my clone has published 750+ videos on, so I’ve got skin in this. I’d just rather you read a server log than a dial, so the rest of this post is me showing you which twelve things on that log are worth your time.
LLM SEO is three steps and you can watch two of them
When ChatGPT, Perplexity, Gemini or Google’s AI Mode answers a question, three things happen in order. The engine searches, usually through Bing or Google. It reads a handful of the pages it found. Then it writes, quoting several of the pages it read. LLM SEO is the work of being one of them.
You can watch step one, because search rank and indexing are public. You can watch step three, because the citations are printed under the answer. Step two is the black box. Every honest LLM SEO claim lives on one side of it or the other.
The method for all three is in my generative engine optimization guide, with the answer-first writing side in answer engine optimization. This post (call it LLM optimization or GEO if you like, it’s the same job) is about which parts of AI search optimization you can check with your own eyes.
The stakes went public on October 29, 2026, when Reddit’s CEO linked a drop in Google search traffic to AI Mode and AI Overviews and the share price lost about a third in a week.
That’s Reddit, the site half of marketing LinkedIn told you to seed threads on all through 2025. If a company that size is still working out what AI search does to its traffic, you can forgive yourself for squinting at a stranger’s dial.
The 12 LLM SEO signals you can measure this week
Every one of these can be read in a tool you already pay for, or in none at all. The table is the whole list. The notes under it are where the numbers come from, so you can argue with the data, which is more fun than arguing with me.
| # | Signal | Where you read it | Minutes a week |
|---|---|---|---|
| 1 | Whether an engine names you at all | A fixed prompt panel, run by hand | About 120 |
| 2 | Which of your pages gets cited | The same panel, logging the URL | 0 extra |
| 3 | Your Google position for the same question | Search Console, Performance, filter by query | 5 |
| 4 | Whether Bing has you indexed | Bing Webmaster Tools, free | 5 |
| 5 | Which AI bots fetch your pages | Server log or CDN bot report | 10 |
| 6 | AI referral sessions | GA4, referrer contains chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com | 5 |
| 7 | Generative AI impressions | Search Console, Performance, Generative AI tab | 5 |
| 8 | Mentions on pages you don’t own | Brand report in any SEO tool, or a quoted Google search | 15 |
| 9 | Heading match to the question | Your H2s read beside the panel questions | 10 |
| 10 | Quotes and statistics on the page | A count, per page | 10 |
| 11 | Length and focus | Word count, per page | 2 |
| 12 | Share of voice over 30 days | The same panel, scored against two competitors | 0 extra |
1. Whether an engine names you at all
Write 20 questions a buyer would ask in your category, run each one three times in ChatGPT, Perplexity, Gemini and Google AI Mode, and log every mention. Three runs isn’t me being fussy. AirOps sent 16,851 queries to ChatGPT three times each for its 2026 fan-out study because single answers wobble. A mention rate across 60 runs per engine is a number you can track. One screenshot is a mood with a timestamp.
2. Which of your pages gets cited
Engines cite URLs, so log the exact page each time. Don’t be surprised when the citation lands on a post you’d half forgotten while the page you’d have chosen sits untouched. That tells you which of your pages already reads like an answer, so that one goes to the top of the update list.
3. Your Google position for the same question
Ahrefs’ March 2026 update found 37.9% of URLs cited in AI Overviews also sit in Google’s top 10 results for the same query, with the rest split almost evenly between positions 11 to 100 and beyond the top 100. In AirOps’ ChatGPT data the top page in ChatGPT’s own search results had a 58% chance of being cited and the page at position 10 had 14%.
Ranking still gets you into the room. In ChatGPT it does most of the work, while in AI Overviews it’s under 40% of the story now.
4. Whether Bing has you indexed
ChatGPT’s search layer leans on Bing. Seer Interactive matched more than 87% of ChatGPT search citations to Bing’s top 20 results for the same question in February 2025, against about 56% for Google. Open Bing Webmaster Tools, check your indexed pages, and fix anything missing. I know, it’s Bing. Humor me.
5. Which AI bots fetch your pages
Your server log or CDN bot report shows AI fetches by HTTP user agent. Check those logs alongside your robots.txt controls; the table includes both. The entries are taken from OpenAI’s own crawler docs and Google’s October 30, 2026 announcement.
| Crawler or robots.txt control | What it decides | If you block it |
|---|---|---|
| OAI-SearchBot | Whether your site can appear in ChatGPT search answers | You vanish from ChatGPT citations, though navigational links can still show |
| ChatGPT-User | Live fetch when a person asks ChatGPT about a page | Robots rules may not apply, since a human asked |
| GPTBot | Training data for OpenAI models | You opt out of training only. Search visibility is untouched |
| Google-Extended (robots.txt control) | Gemini training and grounding | No effect on Search or AI Overviews, which Googlebot handles |
| __usage_ai-training (robots.txt control token) | One robots.txt entry for AI training preferences across providers | Google plans to honor it from early 2027; a disallow is read as Google-Extended disallowed when Google-Extended is not explicitly specified |
Block the wrong line and you’ve opted out of citations while thinking you protected your content. Read your robots.txt this week, before somebody’s plugin edits it for you. If you’d rather read it with someone beside you, that’s a Tuesday session inside First Movers AI Labs.
6. AI referral sessions in your analytics
The numbers will be small. Record them anyway, since the trend is the point and this is the only place a citation turns into a human on your site. OpenAI reported more than 900M weekly ChatGPT users on February 27, 2026, so a thin slice of a big number still deserves a monthly look.
7. Generative AI impressions in Search Console
Google introduced a Search Generative AI performance report on June 3, 2026. It sits on every property now and shows how often your pages appear inside AI Overviews and AI Mode, by page, country and device. It doesn’t show clicks. On November 2, 2026 Gary Illyes gave an answer about future click data that Barry Schwartz read as a polite no.
Impressions are still a real signal. Put them beside the organic clicks for the same URL and you’ll see which pages Google quotes without sending anyone your way.
8. Mentions on pages you don’t own
Ahrefs ran a correlation study across 75,000 brands in May 2025. Branded web mentions correlated with AI Overview visibility at 0.664, while backlinks came in at 0.218. Lists, reviews, podcasts and comparison posts are the pages engines read when they decide who you are.
My own cheapest mention engine was an accident. The 750+ videos my clone has fronted all sit on youtube.com, which Ahrefs now calls the most cited domain in AI Overviews. Every one of them names First Movers AI out loud. Ahrefs’ December 2025 follow-up across the same 75,000 brands found YouTube mentions the strongest signal of AI visibility, at about 0.737 across ChatGPT, AI Mode and AI Overviews. I’d planned a YouTube channel. I hadn’t planned 750 citable pages on the one domain AI Overviews cite most.
9. Heading match to the question
AirOps found pages whose headings closely matched the query were cited 41% of the time, against 29% for weak matches. So read your H2s out loud next to the 20 questions from signal 1. If the heading doesn’t sound like the question, you’ve found your first edit.
10. Quotes and statistics on the page
The 2023 GEO study (Aggarwal et al.) tested nine ways to change a page and measured visibility in generative answers. Adding quotations, adding statistics and citing sources came out on top, with gains of up to 40% on the paper’s own benchmark, while keyword stuffing scored below the untouched page. Count the dated figures and named sources on your page. If the count is zero, there’s your afternoon.
11. Length and focus
AirOps put the citation sweet spot at 500 to 2,000 words, while pages over 5,000 words were cited less often than pages under 500. Pages covering a quarter to a half of ChatGPT’s follow-up sub-queries beat pages trying to cover all of them. Yes, I see the irony. This one runs long because it’s two lists and a scorecard, so skim to the part you came for.
12. Your share of voice over 30 days
Take the same 20 questions, add two named competitors, and run the panel weekly for a month. Ahrefs found AI Overview content changes every 2.15 days on average, with 45.5% of cited sources new from one observation to the next.
Semrush, in its 2024 study, watched a URL hold its place in an AI Overview for under four consecutive days on average. One reading tells you nothing. Four weekly readings give you a line, which is the first thing a dashboard vendor can’t fake.
The LLM visibility scorecard I’d build before paying anyone
Signals 1, 2 and 12 are one spreadsheet, so let me give you the spreadsheet. One tab per week. Rows are your 20 questions. Columns are the four engines, three runs each, which is 240 cells a week. Each cell gets one of four scores.
| Score | What happened in the answer | What it tells you |
|---|---|---|
| 0 | Your brand is absent | Your brand was not surfaced in this answer |
| 1 | Your brand is named, no link | Your brand was named without a citation; its source cannot be determined from the answer alone |
| 2 | One of your pages is cited | You were read and quoted. Log the URL |
| 3 | Your page is the first citation | You are the primary source for that question |
Add the 240 scores and divide by 720, the most you could score, to get a weekly percentage. Do the same for the two competitors in the same cells. By week four you have three lines on one chart, which is the first useful conversation you can have about LLM visibility.
If your team will run the scorecard, use my AI training for employees program to turn it into a weekly habit.
Write the questions the way a buyer talks, which means three shapes. Category questions (“what’s the best AI consulting firm for a 20-person company”), comparison questions (“First Movers AI vs a generic AI agency”) and problem questions (“how do I stop spending 90 hours a week in my own business”). The AI visibility audit post has the 10 I’d run for First Movers AI if you want a worked set.
Now the math on the alternative. $250 a month is $3,000 a year. The scorecard costs about two hours a week at 30 seconds a run, with every one of those 240 cells yours to see. In July 2025 Pew reported on 68,879 real Google searches and found people clicked a result 8% of the time when an AI summary appeared, against 15% without one.
Your clicks are thinning whatever you spend, so spend the $3,000 on the pages that get quoted. Spend the two hours finding out which ones those are.
The 20 LLM SEO claims being sold without the evidence
The problem starts when the invoice arrives before the proof does. Some of these are reasonable hunches that a good vendor will happily show their work on. The table is blunt on purpose. The five that cost people the most money get a longer look underneath.
| # | The claim | What the evidence says | Pay for it? |
|---|---|---|---|
| 1 | An AI visibility score out of 100 | Ask how many prompts, runs, engines and dates sit behind it. One prompt, one run has the shelf life of milk | Only with the appendix |
| 2 | llms.txt | 97% of files on 137,000 domains were never requested by anything in May 2026 (Ahrefs) | No |
| 3 | Prompt volume | No engine publishes it. Every figure is modeled, so it should be labeled an estimate | As a direction only |
| 4 | Guaranteed citations in 90 days | 45.5% of cited sources change between two observations (Ahrefs). A guarantee is a promise about a coin | No |
| 5 | Entity work as a sealed package | Entities are real. With no before-and-after mention count there is no way to tell the work happened | Only with a mention count |
| 6 | Sentiment scores | Wording changes every regeneration. A single-run sentiment reading is noise | Only with run counts |
| 7 | Your position inside the answer | There is no stable position in a generated paragraph. Retrieval rank is the measurable thing | No |
| 8 | Backlinks for AI | Backlinks 0.218 vs plain mentions 0.664 (Ahrefs). No positive correlation in AirOps data | For Google, yes. For AI, no |
| 9 | Domain authority for LLMs | DR 0.326, half the figure for mentions (Ahrefs). AirOps found no positive correlation | No |
| 10 | Schema as the switch | JSON-LD pages cited 38.5% vs 32.0% (AirOps). Worth adding, nowhere near a switch | Add it yourself |
| 11 | Keyword density for LLMs | Keyword stuffing scored below baseline and about 10% worse on Perplexity (GEO study) | No |
| 12 | Markdown mirror pages for bots | Mueller has called separate Markdown pages for bots a stupid idea. Engines read your HTML | No |
| 13 | Blocking GPTBot as protection | GPTBot is training only. OAI-SearchBot governs ChatGPT search inclusion | Read the file yourself |
| 14 | An AI algorithm update calendar | OpenAI and Perplexity publish crawler docs, no ranking docs. The calendar is model releases relabeled | No |
| 15 | Minimum word counts for AI | 500 to 2,000 words was the sweet spot. Over 5,000 cited less than under 500 (AirOps) | No |
| 16 | Reddit seeding as a shortcut | Reddit’s own CEO tied its traffic drop to AI Mode on October 29, 2026. One source is rented space | As one channel, not a plan |
| 17 | One dashboard for every engine | Only 9% of teams track all relevant metrics across platforms; Gemini mention-to-citation overlap as low as 30% (Semrush) | Only if it shows per-engine data |
| 18 | Revenue attribution to AI answers | Referrers go missing, and a ChatGPT reader who Googles you later shows as branded search. It is a model | Only with the model visible |
| 19 | Cracking ChatGPT’s algorithm | There is no published one. Search, read, write. The measurable parts are in the first list | No |
| 20 | The monthly PDF with no prompt list | Without prompts, engines, dates and raw responses the number is a feeling with a logo on it | Ask for the appendix |
llms.txt, the file nothing reads
Ahrefs checked server logs on 137,210 domains in May 2026. Of the roughly 38,000 with an llms.txt file, 97% never had it requested once, by anything. John Mueller said on Google’s own podcast in June 2026 that engines can’t trust a file sites write about themselves to choose between them. Publish one if you like, but don’t pay anyone to write it for you.
The visibility score and the four questions that undress it
How many prompts. How many runs. Which engines. Which dates. Semrush’s own 2026 index found 45% of marketing leaders can’t accurately measure their AI visibility, so the market for a tidy number is enormous. A vendor who answers all four in the first email is usually worth a second one. A vendor who answers with another screenshot has told you what the score is made of.
Guaranteed citations, or a promise about a coin
With 45.5% of cited sources changing between two observations, a guarantee is a promise about which way a coin lands next month. The question to put to them is what happens to the fee when the citation shows up in week three and vanishes in week five. The answer to that question is the contract.
Blocking AI bots, the protection that can cost you the citation
Blocking GPTBot opts you out of training and nothing else, since OAI-SearchBot governs ChatGPT search on its own. A blanket AI-bot block that catches OAI-SearchBot too tends to set up a second invoice for citation recovery down the line. Go back to the bot table in signal 5 and read your robots.txt against it.
Backlinks and authority, the old product with a new label
Ahrefs put backlinks at 0.218 against 0.664 for plain mentions, with Domain Rating at 0.326. AirOps found no positive correlation between domain authority and citation. Links still do their old job in Google. Sold as an AI lever, they’re the old link-building product with a new label that the data above doesn’t support. On this data, a small site with a dated, quotable page can out-cite a big site with a vague one.
Your own logs are the best LLM visibility tool you’ll ever own
Let me ask you this. If you had to choose between a $250 dashboard and about three hours a week across the prompt panel, Search Console, Bing Webmaster Tools and your server log, which one would you defend in front of a client who asks where the number came from? I mean, you already know. The free tools are first-party. The dashboard is somebody else’s sample of somebody else’s sample, with no way for you to see either one.
I came to LLM SEO the slow way. I sold my 100-person agency in 2021 and started again with two people. I passed $100K a month by month 11 of the rebuild.
The public face of that rebuild is a YouTube channel at 300,000+ subscribers as of November 2026, run by a clone that has fronted 750+ videos. The tracking that told me most was whether the next page answered a real question better than the last one. It still is.
The engines reward the same things readers do. Dated numbers and named sources, a heading that sounds like the question, and a page that stops when the answer’s done. Build for that and the 12 signals move on their own, which is also why AI SEO and old-fashioned SEO keep turning out to be the same discipline wearing different badges. The 20 can wait until somebody proves them.
FAQ on LLM SEO
What is LLM SEO?
LLM SEO is the practice of making your pages the ones large language models retrieve and cite when they answer a question in ChatGPT, Gemini, Perplexity or Google AI Mode. It overlaps almost entirely with generative engine optimization and answer engine optimization. The work is search visibility, clear page structure, quotable facts and brand mentions on pages you don't own.
How do I measure LLM visibility without paying for a tool?
LLM visibility can be measured with a fixed panel of 20 buyer questions run three times each in the four main engines and scored 0 to 3 weekly for a month, plus Search Console's generative AI report, Bing Webmaster Tools, AI referrers in GA4 and the AI crawler lines in your server log. Together those give you mention rate, cited URL, impressions, referrals and crawl access without a subscription. The First Movers AI visibility audit walks through the scoring.
Is llms.txt worth adding?
Rarely. llms.txt is a text file listing your important pages for AI tools, and Ahrefs' May 2026 server-log study of 137,210 domains found 97% of published files were never requested by anything. Google's John Mueller has said engines can't rely on a self-written file to choose between sites. It costs nothing to add and should cost nothing to have added for you.
Does ranking on Google still matter for LLM SEO?
Yes. Ranking on Google still matters for LLM SEO because engines retrieve pages through search before they write. In AirOps' 2026 ChatGPT study the top retrieved page was cited 58% of the time and the page at position 10 14%. Ahrefs found 37.9% of AI Overview citations also sat in Google's top 10 for the same query in March 2026. Rank is the entry ticket, with page structure deciding who gets quoted.
What is the difference between LLM SEO and GEO?
LLM SEO and GEO describe the same job from two angles. GEO, generative engine optimization, is the name from the 2023 GEO study (Aggarwal et al.) and covers the whole method. LLM SEO is the phrase people reach for when they mean the SEO parts of it, which are search rank, indexing, crawl access and structure. AEO narrows further to writing the answer itself. The SEO vs GEO post lays the three side by side.
Build the scorecard with my team watching your screen
Say you bring your 20 questions to a live build inside First Movers AI Labs. We run twice a week, Tuesdays and Wednesdays, with 60+ courses behind the sessions and 5,000+ students trained by me and my team. You’d leave the first one with the scorecard tab built and your first engine scored, your robots.txt read against the bot table, and your worst vendor email answered with four questions. Join AI Labs and bring the email.
The dials will keep arriving, so reply to the next one with those four questions. How many prompts, how many runs, which engines, which dates. That’s the whole of LLM SEO due diligence in one line, and the 12 signals are sitting in tools you logged into this morning.
SEO will die again next Tuesday. Your log will still be there on Wednesday.
