LLM SEO: the guide written from measurement

LLM SEO is the practice of making a brand visible inside the answers of AI engines such as ChatGPT, Perplexity and Google AI when buyers ask what to choose. Botanik measures that visibility in three states, Absent, Mentioned and Recommended, factor by factor, with Oracle, the tool built by Maple Labs.

Ask ChatGPT which online bank to choose, and it answers with a shortlist. Your brand is in it, or it is not. Botanik scores that outcome in three states, on each decision factor: Absent (not cited), Mentioned (cited without being presented as the right choice) and Recommended (presented as the option that fits, or the preferable one). One question typed once proves nothing. The states come from more than 2,500 AI searches analyzed in a single audit, and that is the point.

This guide takes one position: measurement before opinion. Every number on this page comes from a published source: chiefly the Baromètre GEO France, edition T3 2026, collected on August 1, 2026 and reusable under CC BY 4.0, plus the Méthode Oracle page (method by Maple Labs), the public price grid and the buying guide updated July 2026. Nothing here is a screenshot, an anecdote, or a vibe.

One conviction frames the work, stated in Botanik's manifesto: visibility is cultivated, it is not hacked. There is no trick that flips an engine's answer overnight. There is a measurable state, a set of levers, and a re-measurement that tells you whether the levers worked. That loop is what the rest of this page describes.

AI engines do not share one ranking. ChatGPT leans on its corpus and on Bing search, while Gemini draws on the Google index and its SEO signals. In Botanik's Baromètre GEO France, the 3 engines tested named a different number 1 in 2 of the 4 published markets.

Buyers question engines with four intents: discovery ("how do I choose…"), comparison ("which is the best…"), local ("… near me") and validation ("reviews of…"). Oracle does not turn them into a prompt list. Replica agents play prospects and move the search forward like real buyers, and the topics come up on their own in the conversations.

Each engine assembles its answers from different raw material. ChatGPT relies heavily on its corpus and on Bing search. Gemini draws on the Google index and its SEO signals. The same brand can come first on one engine and lower on another, which is why Oracle works on ChatGPT, Claude and Gemini and builds into its analysis any difference between assistants that would change the conclusions.

Answers also move between runs of the same prompt. That variability is not a bug you can wish away. It is a property you measure. The findings below quantify it.

Citability is earned in two places: on your own pages and in the sources engines read. Botanik's implementation work targets three recurring technical barriers, heavy JavaScript, a missing llms.txt file and incomplete Schema.org markup, then builds presence in the sources engines cite, from comparison sites to Wikipedia and Reddit.

On the page, two things matter in Botanik's implementation work: content designed to be cited, and the removal of three recurring technical barriers: heavy JavaScript, a missing llms.txt file, and incomplete Schema.org markup. Adoption of llms.txt is still marginal: 3.2% of sites have deployed one. The field is early, which means most of your competitors have not moved yet.

Off the page, engines cite sources. Botanik places clients in the sources engines actually draw from: comparison sites, press and regional dailies, authority directories, Wikipedia, Reddit and LinkedIn. This is slower work than publishing on your own domain, and it is where a large part of citation is decided: an engine cannot recommend what its sources never mention.

Which sources matter most depends on the factor where you are losing. ChatGPT leans on Bing and its corpus. Gemini leans on the Google index and its SEO signals. Oracle shows, factor by factor, which coverage sources the AI relies on, and that is where to put the effort first.

First citations in LLM SEO move in 6 to 10 weeks. A durable shift from Mentioned to Recommended plays out over 4 to 6 months, because engines, classic and AI alike, take 3 to 4 months to re-crawl, re-cite and re-rank. Botanik's Pousse plan (Pousse, the sprout) carries a 6-month engagement for that reason.

Botanik's buying guide sets a hard rule for any provider: 30 days is the deadline by which you must have a deliverable in hand, not a discovery phase. Botanik's own calendar works on that clock. At 30 days: your status, Absent, Mentioned or Recommended, factor by factor, a baseline signed and measured by a third-party tool, and a prioritized roadmap. At 60 days: content and structured data published, placement under way in the sources engines cite, and a first dated competitive comparison. At 90 days: a report against the baseline, your movement against the competitors the AI compares you with, and a decision: continue, adjust, or stop.

The 6-month engagement on Pousse follows from engine behavior, not from sales preference. Engines, classic and AI alike, take 3 to 4 months to re-crawl, re-cite and re-rank a market. Below 6 months, you pay to sow without ever harvesting.

Botanik's free scan answers one question in 60 seconds: do AI engines recommend you? It needs no account and no card and returns a verdict per engine: Absent, Mentioned or Recommended.

The scan queries ChatGPT with and without web search, Gemini and Perplexity: the four surfaces it tests. Pick your sector, get your verdict per engine. It is the fastest honest answer to "where do I stand", and it costs you one minute.

If you would rather understand the method first, our method page lays out Maple Labs' 9 steps: framing, conversations simulated at scale, decision factors, trusted sources, consolidation, one activation prompt per factor, identifying the recommended players, tested until stable (variance under 5%), coverage brand × factor × source, then scoring and prioritization.

Two rules keep the numbers honest. Statuses: Absent means not cited on a decision factor, Mentioned means cited without being presented as the right choice, Recommended means presented as the option that fits, or the preferable one. Stability: each activation prompt is tested until the result is stable, with variance under 5%, before anything is scored. A single answer is never treated as a result.

Cite it or rebuild it: both are allowed for the Baromètre GEO France, free to reuse with attribution (CC BY 4.0). Its scoring follows its own frozen protocol, published on the Baromètre page; it is a Botanik study, separate from Oracle. The Baromètre keeps every collected line as JSON: exact prompt, verbatim response, model actually served, timestamp, attempts, tokens consumed. The full trace set is available on request, exactly as written, unedited.

Honest LLM SEO states its limits. Maple Labs lists five for Oracle, and we repeat them upfront: it does not see the prompts your users actually type, does not guarantee a recommendation, does not manipulate answers, does not replace SEO, and does not predict every answer.

A visibility score is also silent on the things that matter most to your P&L. The Baromètre states it plainly: it says nothing about your revenue, your market share, or what an AI will answer to a prompt outside the measured list. A real user's answer also depends on their account, their history and their country, none of which a panel fully reproduces.

Honesty is operational, not rhetorical. When Perplexity's API key was not active in the Baromètre's collection environment, the engine was not queried and no value was extrapolated in its place: the gap is published as a gap. The protocol was frozen on August 1, 2026 at 08:03 UTC, before the first call, with a published SHA-256 fingerprint. Hold every provider, Botanik included, to that standard.

The same discipline applies to guarantees. A "+30% guaranteed" with no contractual indicator, no signed baseline and no measurement window is a commercial promise, not a guarantee. Botanik's own local guarantee (+30% local visibility on your zone, or the variable share of its fees is not due) stands on three written conditions: a contractual indicator, measurement by a third-party tool, and a 1-month measurement window. As Botanik's manifesto puts it: a number without a method is an opinion.