How to appear in Meta AI
The engine inside WhatsApp and Instagram exposes neither an API nor its sources. What is known, little as it is, the two Meta agents, the right name on the pages it reads, proximity questions, and how you measure it in a real session when there is no other way.
Diego MoraesCTO
Open in ChatGPT
Open in ClaudeMeta AI is the engine most Portuguese people have at hand, inside WhatsApp and Instagram, and the one that lets itself be measured least: no API, no visible sources. This guide says the little that is known and how the rest is measured.
How Meta AI decides
Through Llama, with its own search. It answers from what Meta indexes of the open web. It exposes neither an API nor sources: what gets measured is what a person sees in a real session.
Through the name in the text. With no visible citations, the signal that counts is the brand said with the right name in the answer, and the absence is the data. What appears comes from the pages Meta reads: the brand's own, the directory listings, Wikipedia where it exists, and the press.
Close to the buyer. WhatsApp questions are quick, local decisions. In our measurements, the answers carry places from the map and what Meta found about each one.
The absence is the data.
What to do, in order
Let Meta read the site
Meta AI answers from Llama with its own search over what Meta indexes of the open web. The crawler is meta-externalagent, and meta-externalfetcher goes to fetch a page when somebody asks for it in a conversation. Without a way in, the brand exists only through what third parties say about it.
Do: in robots.txt, allow meta-externalagent and meta-externalfetcher. Server-rendered HTML: what only exists after the JavaScript runs is not read.
Have the right name on the pages Meta reads
With no visible citations, the signal that counts is the brand name said in the text of the answer, and the absence is the data. What appears comes from what Meta indexes: the brand's own page, the directory listings, Wikipedia where it exists, and the press. A name written two ways is half a presence.
Do: the same name, in the same spelling, on the site, in the sector directories, on LinkedIn and in the press. If the brand has a Wikipedia article, keep it correct; if it does not, do not force one.
Answer proximity questions, because that is where Meta AI lives
Meta AI sits inside WhatsApp and Instagram, and the questions asked there are quick and local: which clinic, which restaurant, which shop, which service near me. In our real-session measurements, the answers carry places from the map and what Meta found about each one.
Do: one page per service and per location, with the answer to the local question in the first sentence, exact address and opening hours, and the Google listing and the directories saying the same thing.
Measure in a real session, because there is no other way
There is no public Meta AI API and the sources cannot be seen. The only honest measurement is a person, or a system doing what a person does, asking the question in the app, with a dedicated account, from the buyer's country. That is how we have measured Meta AI since 27 August 2026: the same questions as the weekly audit, in a real session, from Lisbon, and the answer kept in full.
Do: if somebody sells you Meta AI measurement by API, ask which API. Measure in the app, every week, and keep the text: with no sources, the text is the only proof.
The check, in one command
What is verified with a command. Swap the domain for yours.
The Meta agents have a way in
$ curl -s https://example.com/robots.txt | grep -iA1 'meta-externalagent\|meta-externalfetcher' User-Agent: meta-externalagent Allow: / -- User-Agent: meta-externalfetcher Allow: /
If there is no block at all for them, the general rules apply; if there is a generic Disallow, they are left out. The full check is in six commands to find out whether the AI can read your site.
What is not known
The pace of Meta's index, the weight of each source, and whether the answer changes between WhatsApp, Instagram and meta.ai. There is no public study that answers this, and we do not invent one. What we have is the series: the same questions, every week, in a real session, and the answer kept in full so it can be compared.
Frequently asked questions
No. It exposes neither an API nor sources. What is measured is what a person sees in a real session: the answer in text, the places from the map when the question is local, and the brand name inside the text or the absence of it. The absence is the data.
From Llama, with its own search over what Meta indexes of the open web: the brand's page, the directories, Wikipedia where it exists, and the press. Appearing means having the brand said with the right name on those pages.
For local and consumer brands, yes: it sits inside WhatsApp and Instagram, where the quick decision happens. For specialised B2B, the weight is smaller and gets measured before you invest. The groundwork (consistent name, legible pages, correct directories) is the same as for the other engines; the marginal cost is small.
There is no public data on the pace of Meta's index, and we do not invent one. What is known from measurement is that the answer changes from week to week, and that is why it is measured every week instead of a deadline being promised.
Read next
- Knowledge vs augmented, the same rule of the real session, on the other engines.
- How to appear in ChatGPT, the engine with the most public measurement.
- How to appear in Claude, the other engine measured in a real session.