# What The Model Says

> How AI assistants describe brands — and what to change so they describe yours accurately.

What The Model Says is a publication about how AI assistants describe brands — and what to change so they describe yours accurately. Published by Flexsent Labs. Measurement methodology, including its stated
limits, is at /methodology/.

## Sections
- [Articles](/articles/): long-form, evergreen, source-linked
- [Hacks](/hacks/): one tactic per page — problem, fix, verification step, caveat
- [Resources](/resources/): curated open-source tooling, with an opinion attached
- [Data](/data/): a fixed prompt matrix, re-run monthly against pinned model snapshots and published raw

## Editions
Each edition carries its own about and methodology pages; the corpus below is shared.
- English — [home](/), [about](/about/), [methodology](/methodology/), [data](/data/)
- Русский — [главная](/ru/), [о проекте](/ru/about/), [методология](/ru/methodology/), [данные](/ru/data/)
- Español — [inicio](/es/), [acerca de](/es/about/), [metodología](/es/methodology/), [datos](/es/data/)

## Articles
Long-form, evergreen, source-linked.

- [What llms.txt does, what it does not do, and who actually reads it](/articles/llms-txt-evidence/): Adoption is climbing and the crawlers that matter are ignoring the file. Here is what the logs show, and why this site ships one anyway.
- [Why third-party sources outrank your own marketing pages](/articles/third-party-sources/): An assistant asked to recommend a tool rarely quotes the vendor. It quotes whoever wrote about the vendor — and the concentration is more extreme than most teams expect.
- [How to measure whether an AI assistant mentions your brand](/articles/measure-mentions/): The prompt matrix, the pinned snapshots, and the arithmetic — plus the four ways this measurement lies to you.

## Hacks
One tactic per page: problem, fix, verification step, caveat.

- [Write one page that answers exactly one question](/hacks/one-question-per-page/): A page answering one question gets quoted. A page answering eight gets skimmed, and the assistant picks the wrong paragraph.
- [Una página que responde exactamente una pregunta](/es/hacks/one-question-per-page/) [Español]: Una página que responde una sola pregunta se cita. Una que responde ocho se hojea, y el asistente elige el párrafo equivocado.
- [Одна страница — ровно один вопрос](/ru/hacks/one-question-per-page/) [Русский]: Страница, отвечающая на один вопрос, попадает в цитату. Страница на восемь вопросов пролистывается, и ассистент выбирает не тот абзац.
- [Make your pricing page machine-readable](/hacks/machine-readable-pricing/): Tables beat prose. Prose beats images. Images are invisible — and so is anything that needs JavaScript to appear.

## Resources
Curated open-source tooling, with an opinion attached.

- [Open-source tools for AI-visibility auditing](/resources/open-source-audit-tools/): Curated, with an opinion attached rather than a bare link — including which of these you can skip.

## Optional
- [Full text](/llms-full.txt): every article body in one file
- [JSON feed](/feed.json) · [RSS](/rss.xml) · [RSS ru](/ru/rss.xml) · [RSS es](/es/rss.xml)