LLM SEO
LLM SEO: the playbook, minus the folklore
LLM SEO is the work of making a brand something large language models name, and describe correctly, when people ask them buying questions. There is no ranking to climb: the answer is regenerated per question, per model, per day. What you can change are its inputs, which is what this playbook is for.
People arrive at this topic asking how to rank in ChatGPT, and that framing is the first thing to unlearn. Nothing here is a position you hold; it is a description the model reassembles every time. Every step below carries the number, from our own 5,190 recorded answers, that earns its place on the list.
Updated 3 Nov 2026 with our 18 Oct 2026 agent benchmark, the buyer-question method and a measurement plan.
STEP 1
Measure before you touch anything
Ask the models the questions your buyers type, and keep the transcripts. Not a score: the actual answers, with who was named and what was cited. Everything after this step is spending; this step is the only thing that tells you whether the spending worked.
Observed
Our free check does this across 7 models and keeps every word. Run it on your domain before reading the rest of this list.
STEP 2
Cover every model, because one model tells you almost nothing
The single most repeated mistake in this field is optimising for ChatGPT and assuming the rest follows.
Observed
Across 718 questions answered by more than one model, the average overlap between two models’ brand lists was 24.5%, and of the ~28.3 brands named per question, under one was named by all models.
STEP 3
Get described where the engines already cite
The answer is assembled from pages you do not own. The work is being accurately described on the roundups, comparisons, reviews and communities the engines already reach for in your category, and your own site is the one source that cannot vouch for you.
Observed
Citations in our corpus spread across 17,829 distinct domains, the top 20 holding just 6%. Breadth beats a short list.
STEP 4
Take review platforms as seriously as Reddit
The folklore says Reddit and Wikipedia are the whole game. Live citations disagree: comparison and review platforms appear more often than either, because they are shaped exactly like the answer the engine is trying to write.
Observed
Share of our answers citing each at least once: review platforms 9%, Reddit 4%, Wikipedia 2%. Wikipedia’s influence is real but lives in the memory layer, not in live retrieval.
STEP 5
Write pages an engine can lift
Answer in the first forty words, make headings questions, put a number or a definition where a quote would go. Extraction happens at passage level, and the passage that gets lifted is the one that stands alone.
Observed
The full version of this is its own guide: writing the page that gets lifted.
STEP 6
Feed the memory layer, not just retrieval
A meaningful share of answers cite nothing at all: the model answers from what it already believes about your category. You do not reach that layer with one page. You reach it by being described consistently, in the words buyers use, across enough of the web for long enough.
Observed
19% of all recorded answers exposed zero sources. For those, whatever the model already believed was the whole answer.
STEP 7
Re-run the same questions, on a schedule
The answers are regenerated continuously, so a single measurement is a snapshot. Keep the question set fixed, re-run it, and compare transcripts. Movement you can see beats vibes from a dashboard, and it is the only honest way to attribute results to the work.
How AI assistants choose sources: what we measured
The steps above say where to work. This is the evidence, from 2 dated measurements of our own categories. Assistants search, read a handful of pages and write 1 answer, and the pages they read are rarely the ones they name.
13.8%
of answers cited an essentras.com page
48 of 349 answers
7.7%
named Essentras in the text
27 of 349 answers
4.9%
picked Essentras first
17 of 349 answers
Read is not recommended. Assistants cited our pages almost twice as often as they named us. Our own tracker shows the same gap: on 10 press release prompts, 7 assistants and 3 weekly runs (19 Oct to 2 Nov 2026), 0 of 210 answers named Essentras and 20 cited essentras.com.
Assistants quote the rule first. The page cited most was Google's help center, support.google.com, in 53 answers. Apart from it and Wikipedia, the other 8 of the 10 third-party domains cited most belong to companies selling in these markets, writing about their own field.
Some assistants rarely show sources: from 100% of answers (Perplexity, DeepSeek, Grok) and 98% (ChatGPT) down to 68% (Meta AI) and 44% (Gemini). Those can still name you; you just cannot see what they read.
| Domain | Answers | What it is |
|---|---|---|
| support.google.com | 53 | Google's help center: the platform's rules |
| instantpress.co | 34 | Placement agency |
| einpresswire.com | 27 | Press release distribution |
| reputationresolutions.com | 25 | Reputation management |
| ereleases.com | 20 | Press release distribution |
| badenbower.com | 18 | PR agency |
| reviewtactic.com | 18 | Review management |
| en.wikipedia.org | 17 | Encyclopedia |
| redpress.net | 17 | Press release distribution |
| kalicube.com | 14 | Knowledge Panel agency |
In our audit of 3 Nov 2026 we sorted 92 sources that shape the answers to our buyer questions, found in 559 assistant answers and in searches for the same questions, by what a brand can do about each: 66 had to be earned, 14 were profiles to claim, 6 were owned pages and 6 carried wrong or outdated information. The gap is mostly off your site.
Source mixes move, too. Discovered Labs reported on 20 Oct 2026, citing the tracker Promptwatch, that Reddit's share of ChatGPT Search citations fell from 3.83% (18 Jul to 7 Aug 2026) to 0.52% (14 to 17 Aug 2026), a drop Promptwatch calls provisional; opinion questions still pull Reddit in, while comparison and pricing questions read vendor pages.
What it changes in the playbook
- State the platform's rule in its own words and link it: assistants quote it first.
- Work on the lists, comparisons and reviews assistants read before they name vendors. Your own page gets read; it rarely gets you named.
- Check which assistants show sources at all before you judge a citation rate.
The buyer-question method: from 1 question to 1 change
LLM SEO fails when it starts from keywords and ends at a visibility score. This method starts from a buyer's decision and ends at a paid order. It follows the process Discovered Labs published in its case study of Granola, the AI notetaker, and we applied it to 149 buyer questions of our own on 3 Nov 2026.
01
Write the buying decisions down
From sales calls, support tickets and search data: 1 row per decision, in the buyer's words.
02
Give each question 3 forms
The Google keyword, the AI prompt with its scenario or constraint, and the 1 page that should answer it. 1 intent, 1 URL.
03
Measure both sides
Google: impressions, position and clicks of that page. AI: who each assistant named and cited.
04
Diagnose before you write
Give every lost question 1 of the 4 causes below. Only 1 cause is fixed by a new page.
05
Pick a route for each source
For the exact page an assistant reads instead of yours: OWN, CLAIM, EARN or CONTEST.
06
Make 1 change, then ask again
Re-run the same questions after the change. A fixed prompt set is what makes the comparison mean something.
07
Judge by orders, not visibility
Granola's case counted activated users. For us, it is paid orders whose first touch was an AI assistant.
| Cause | What it looks like | The fix | Ours |
|---|---|---|---|
| Not seen | The page exists, but Google rarely shows it and assistants do not cite it | Technical: indexing, internal links, crawl access | 75 |
| Outranked by third parties | Your page is read or shown, but other brands are named | Off-site: the lists, reviews and profiles assistants read | 53 |
| Weak page | The page does not answer the question asked | Rewrite that page instead of adding another | 15 |
| No page | Nothing on your site answers it | Write 1 page, or a section on a page that ranks | 6 |
- 143 of our 149 questions already had a page. The work was being seen and corroborated, not writing more.
OWN, CLAIM, EARN, CONTEST: 1 route per source
Discovered Labs' matrix sorts each source by your relationship to it; a different relationship needs a different move. A paid placement sits outside it: it is paid, never earned.
| Route | Your relationship | The move | Examples | Where Essentras can help |
|---|---|---|---|---|
| OWN | You publish and edit it | Answer first, dated facts, 1 page per intent | Service, pricing and comparison pages | Free AI check; AI SEO service on a scoped quote |
| CLAIM | A profile you can take over under the platform's rules | Claim it, keep it complete and current | Business profiles, Knowledge Panel, Wikidata, review sites, app stores | Knowledge Panel, $650 individual or $1,200 business |
| EARN | An independent owner decides what it says | Give the owner something worth using: data, access, a proven correction | Coverage, independent lists, reviews, community threads | Nobody can sell earned coverage, us included |
| CONTEST | It is wrong or outdated, or shows only competitors | Correct it through the owner or an official channel, or ask to be included | A stale listing, a wrong fact, a review that breaks policy | Review removal from $499 each; name removal from Google from $1,650; official routes only |
| PAID (outside) | You pay for the slot and approve the text | Buy it as what it is and call it paid | Sponsored articles, press releases | Named outlets at catalog prices, each label shown |
The Granola case, read at the source
Discovered Labs' case study (read 3 Nov 2026) describes 1 plan for Google and AI search: buyer decisions mapped to keywords and prompts, every gap sorted into content, technical or third-party work, the 2 measured separately, and success counted in activated users.
It reports AI-attributed monthly activated users up 4,600% since January 2026 (47 times January), and non-branded AI citation and mention rates up 558% and 739% from January to August 2026, measured by Granola's attribution and the agency's own tool, with no baseline published. A percentage without a base says little: by our arithmetic, a mention rate that rose 739% cannot have started above 11.9%. Copy the method, not the target.
Where LLM SEO budgets actually die
The playbook above is what to do. Just as important, because this market is young and loud, is what quietly eats budgets without moving a single answer:
Buying a dashboard instead of a change
Monitoring tells you the score. It does not play the game. A tracking subscription with no budget left for changing what the models read is a thermometer purchased instead of medicine, and it is the single most common shape of wasted spend we see.
Rewriting your own site for the fifth time
Your site earns you retrievability and correct self-description, and then its returns collapse, because the answer is assembled from other people’s pages. The fifth homepage rewrite competes with the first for the same single slot: witness number zero, the one the engine trusts least about you.
Paying for guaranteed citations
Nobody can sell a position inside a model. There is no ranking, no submission form, no ad unit in the answer. What is honestly purchasable is description: real articles on real sites the engines read. If the invoice says “guaranteed AI citation”, the product is the invoice.
Optimising for last quarter’s folklore
llms.txt files, keyword-stuffed FAQ farms, schema as a silver bullet: each had a season of conference slides, none shows up as a cause in our citations. The habit that survives every season is duller: be described accurately, in many places, and measure.
How to run the measurement yourself
You do not need us for step 1, and this guide is more useful if it shows the whole method. The manual version:
- 1. Write down 10 questions your buyers would actually type, in their words. Pull them from sales calls, support tickets and search data, not from your positioning document.
- 2. Ask each question in a fresh chat, on each assistant your buyers use, with browsing on where the option exists. No fresh chat, no valid reading: history contaminates the answer.
- 3. Record four things per answer: were you named, at what position among the brands, who else was named, and which sources the engine exposed. A spreadsheet is fine. Screenshots age badly; text survives.
- 4. Repeat monthly, same questions, and only compare like with like: the same question on the same engine. Cross-engine averages destroy the only signal you have.
That is exactly what our free check automates, across 7 engines at once, with the transcripts kept for you. The method is not the secret; running it consistently is.
Measure Google and AI separately
A page can climb in Google and not move in ChatGPT, and the reverse. Keep 3 scoreboards that never share a number, and call a change real only when the business scoreboard moves.
| Scoreboard | What to track | Source | How often |
|---|---|---|---|
| Impressions, position and clicks of the owning page; whether it is indexed | Search Console | Weekly; indexing every 3 weeks | |
| AI | Mention and citation rate per assistant on fixed, non-branded prompts; who else is named | Your own runs or a tracker, plus checks by hand in the apps | Weekly for priority questions, monthly for all |
| Business | Visits, free tool use, leads and paid orders by first touch: organic search or AI assistant | Analytics and order records | Monthly |
The terms, defined
Mention rate
The share of answers in a fixed prompt set that name the brand, with or without a link.
Citation rate
The share of answers that list a page from the brand's domain among their sources.
Top pick
The brand an answer recommends first: an explicit pick, or else the first brand named.
Non-branded prompt
A question without your name in it. Branded prompts inflate every number.
Rules that keep the numbers honest
- 1 change per experiment, and the same questions before and after it.
- Run each prompt more than once. SparkToro (28 Jan 2026; 600 volunteers, 12 prompts, 2,961 runs) found under a 1 in 100 chance that ChatGPT or Google's AI repeats the same brand list in 100 runs; presence across many runs was steadier.
- Compare months, not single answers: with 1 answer per task and assistant, a cell under about 20 answers can move 20 points between identical runs.
- API answers are not what each consumer app shows. Check a few by hand in the apps your buyers use.
- Count what pays. Visibility that never reaches a lead or an order is a cost.
Our own baseline, dated
| Scoreboard | Measure | Baseline |
|---|---|---|
| AI | Named, 50 buying tasks | 27 of 349 answers (7.7%) |
| AI | Cited, 50 buying tasks | 48 of 349 answers (13.8%) |
| AI | Named, 10 press release prompts | 0 of 210 answers |
Tools for the AI scoreboard
Free AI check
Asks 7 assistants your buyers' questions and keeps every answer and source.
Run the free checkAI Visibility Baseline
$97 once: what AI assistants say about you, then the first layer to fix.
See the baselineAI Visibility Tracker
$49 a month, $29 per extra domain: the same questions every week, per assistant.
See the trackerSources for the sections added on 3 Nov 2026, each read that day
The research behind these numbers: what the GEO studies and our benchmark show
The sources, category by category
“Get described where the engines already cite” is only useful if you know where that is, and it is different in every category. These are the sites the engines cited most in the categories we have measured deepest, with the checked company’s own domain excluded. This is the table the generic advice cannot print, because it changes per market.
temporary phone numbers SMS verification
714 recorded answers
- burnerapp.com×124
- textverified.com×119
- twoline.io×118
- bestforandroid.com×104
- essentras.com×92
- verifysms.app×88
music dance arts school lessons
322 recorded answers
- superprof.com×52
- thumbtack.com×46
- bark.com×32
- schoolofrock.com×29
- musicteacher.com×24
- coursehorse.com×20
press release distribution and media placement marketplace
301 recorded answers
- fitsmallbusiness.com×103
- prezly.com×91
- ereleases.com×89
- einpresswire.com×81
- pressranger.com×76
- prnewswire.com×69
social media photo and video sharing platform
126 recorded answers
- buffer.com×28
- blog.hubspot.com×17
- futuretechexpert.com×15
- barnimages.com×14
- lifewire.com×14
- apps.apple.com×12
Not on the list for your category? That is what the free check finds: the exact sites the engines quoted when asked about businesses like yours.
The questions with real demand behind them
LLM SEO starts from what people actually ask, so here are the buyer questions with the most measured search volume in our corpus, each linking to what every model answered:
“Which social media platform should I use to share photos and videos with friends and followers?”
30,400,000 searches a month behind it · 7 models answered · names Instagram most
“Which platform should I use to check the Dow Jones Industrial Average today?”
6,120,000 searches a month behind it · 7 models answered · names Yahoo Finance most
“Which company should I use to buy a printable world map poster for my home?”
1,830,000 searches a month behind it · 7 models answered · names Etsy most
“Where can I buy a high-quality world map print for my home?”
1,830,000 searches a month behind it · 7 models answered · names Etsy most
“Which car rental company should I use for my trip?”
1,220,000 searches a month behind it · 7 models answered · names Budget most
“What are the best websites for finding dinner ideas and recipes?”
823,000 searches a month behind it · 7 models answered · names Food Network most
Step 1 takes about a minute
Put your domain in. We pull the questions your buyers actually search, ask 7 live models, and hand you the transcripts: your name highlighted where it appears, the competitors named where it does not.
Run the free checkFree, no account, no card.
FAQ
Frequently asked questions
What is LLM SEO?▾
LLM SEO is the work of making a brand something large language models name and describe correctly when people ask them buying questions. The name emphasises the thing doing the reading: the model. The same work goes by GEO, AEO and AI SEO.
How do I rank in ChatGPT?▾
You do not, because there is no ranking. An assistant regenerates its answer per question, per model, per day, from what it can retrieve plus what it already believes. What you can do is change those inputs: be described accurately on the pages it reads, across every model rather than one, and measure whether the answers move.
Does LLM SEO replace normal SEO?▾
No. The pages that get you cited still have to be found, crawled and trusted, which is classic SEO. LLM SEO adds a second audience for the same work: systems that read widely and answer once. Most of what earns citations, clear pages with something specific on them, described consistently across the web, is good SEO anyway.
How do I measure LLM visibility?▾
Ask the models your buyers’ questions and record the answers. Averages hide everything here, so keep transcripts: which model, which question, who was named, in what order, citing what. That is exactly what our free check does, across 7 models, with every answer kept word for word.
How often do the models change their answers?▾
Often enough that a single check is a snapshot, not a verdict. Retrieval-heavy engines shift with the web; memory-heavy ones shift with model updates. Re-running the same questions on a schedule is the only way to tell a trend from an accident.
What content works best for LLM SEO?▾
Content shaped like the answer the model is about to write. Comparisons, direct recommendations for a named situation, and pages carrying an original number all keep appearing in our citations. Brand essays about yourself almost never do, because the model is looking for a witness, not a self-introduction.
Do I need different content per model?▾
No. You need broader description, not model-specific pages. The models disagree because they retrieve and weigh sources differently, and the same accurate description spread across more of the web is what narrows the gap. Writing "for Claude" or "for ChatGPT" is not a real thing, whatever the vendor deck says.
Is LLM SEO worth it for small companies?▾
It is most valuable exactly where you cannot win classic SEO head-on. An assistant answering a specific buying question reaches for whoever is described well for that question, not whoever has the biggest domain, and our transcripts regularly show small companies named next to giants for precisely phrased needs.
How do AI assistants decide which sources to cite?▾
They search, read a handful of pages and write 1 answer from them. In our benchmark of 18 Oct 2026 the page cited most was Google's help center, then vendors' and agencies' guides; assistants cited our pages in 13.8% of answers but named us in 7.7%.
What is the difference between an AI citation and an AI mention?▾
A mention is your name in the answer text; a citation is a link to your page among its sources. They move separately: in 210 answers to 10 press release questions in our tracker (19 Oct to 2 Nov 2026), essentras.com was cited 20 times and Essentras named 0 times.
What is the OWN, CLAIM, EARN, CONTEST matrix?▾
A way to pick 1 action per source an assistant reads: OWN is what you publish, CLAIM a profile you can take over, EARN independent coverage you must merit, CONTEST wrong or outdated information to correct. A paid placement sits outside it, because paid is not earned.
Should Google and AI visibility be reported together?▾
Side by side, never blended: a page can gain in Google with no change in AI answers, and the reverse. Judge the work by a third number, leads or paid orders split by first touch.