Our methodology is the set of rules we follow to research how AI search engines retrieve and cite pages, and it exists so that you can judge how far to trust any sentence on this site. This page explains what we collect, how we grade it, what we refuse to publish and how we fix mistakes.
The short version
Every guide is built from three kinds of material. We never present the third kind as if it were the first.
| Evidence level | What it is | How we treat it |
|---|---|---|
| Documented | The engine's own help pages, developer docs, crawler lists, official announcements | Stated as fact, with a link and the date we read it |
| Observed | Hands-on tests recorded on video, independent studies with a described method | Stated with attribution, for example "in the walkthroughs we reviewed" |
| Claimed | Agency blogs, vendor studies, forum posts, single case studies | Marked "reported" or "according to", and left out when we cannot find support |
Step 1: collect sources
For each engine we start by gathering everything we can find before writing a line. That means official documentation, saved copies of articles from agencies and publishers, and transcripts of video walkthroughs where practitioners test an engine live or explain what they saw in their own analytics. We keep these in a research archive, so a claim can be traced back to the page or transcript it came from.
Where an engine has little public material, we say so. Mistral Le Chat, Kimi, Meta AI, Qwen, YandexGPT and Brave had no video test transcripts in our archive, so their guides rely on documentation and published analysis and carry a note in the "Limits" section.
Step 2: extract and tag facts
We read every source in full and write down each fact together with its origin. A fact is tagged by where it came from: documentation, a named third-party source, a video transcript, or an archive page. A fact without an origin does not go in. If two sources disagree, we keep both and show the disagreement. Several guides do this, for example when a vendor and the official documentation give different rate limits for the same product.
Step 3: verify against primary sources
Wherever possible we open the primary source behind a secondary claim. A statistic quoted by an agency is only repeated as a fact if we can find the study. If we cannot, the guide says "reported by" and names the agency. Some pages could not be opened at the time we checked, and a few figures stay unconfirmed. Those are listed in each guide's Limits section instead of being smoothed over.
Step 4: write for people, then check the structure
We write for the person who has to make a decision: should I allow this crawler, rewrite this page, ignore this engine? The guides lead with the answer, explain the reasoning, and finish with limits and sources. Each article includes tables where a comparison helps, lists where steps matter, and at least one example you can adapt. Illustrative examples are labelled as illustrations. We do not invent customers, case studies or results.
What we refuse to publish
- Invented statistics, studies, quotes or test results.
- Ranking "weights" for AI engines presented as established fact when only a vendor has published them.
- Claims that we personally ran a test when the finding comes from someone else's video or article.
- Advice that depends on deception, such as fake reviews, hidden text or manipulating a model with invisible instructions.
About the video transcripts
Some guides contain a section called "Analysis of the reviewed walkthroughs". It summarises what the transcripts in our archive actually show. These are third-party walkthroughs, and their claims are self-reported: a presenter's traffic numbers are not independently audited, and a demo on one query does not prove a general rule. We report what was shown, say who showed it in general terms, and avoid turning a single demo into a law.
Updating and corrections
AI products change quickly. Each article shows a published date and an updated date. We revisit guides when an engine changes its documentation, renames a product, or when new evidence contradicts something we wrote. A guide that is out of date is worse than no guide, so when a fact has an expiry we say "as of" followed by the date.
If you find an error, contact us with the page and, if you can, a source. We correct verified mistakes, update the date, and where the change is material we note what changed.
Independence
SEO for LLM does not sell ranking services, does not take payment for coverage, and does not use affiliate links in its guides. The site may display advertising to cover costs. Advertising never influences what we write about an engine or a vendor. You can read how advertising and cookies are handled in our privacy policy.
Who is responsible
The research and editing are done by Mohamed CISSE. Questions about method are welcome.