Meta AI SEO: getting cited inside WhatsApp, Instagram and Facebook

What to know first

1. 1

   meta-webindexer is the one Meta crawler tied to citations: allowing it “helps us cite and link to your content,” per Meta. meta-externalagent collects training data, and blocking it is a separate decision.

2. 2

   Meta named Bing in 2023 and Google in 2024 as search partners. Its AI terms, effective May 13, 2026, say only “select partners, like search engines” — ranking in both is the hedge.

3. 3

   Meta AI’s tool list includes a semantic search over Facebook, Instagram and Threads posts from January 1, 2025, built from captions, visual analysis and transcripts. Say your product’s name, show it and caption it.

4. 4

   Meta documents one source pattern — a Sources list under the response — and no layout per app. In a WhatsApp thread, being named in the sentence does work a link can’t.

5. 5

   There’s no Meta AI console, and GA4’s AI Assistant channel doesn’t list Meta AI. Expect meta.ai referrals from the website, Facebook link-shim referrals from the app, and a lot of Direct.

01Meta AI answers inside chats, not on a results page

Meta AI is one assistant behind many doors: its own app and website, Facebook, Instagram, Messenger, WhatsApp and Meta’s glasses. The busiest door is a messenger. In July 2026 Mark Zuckerberg called WhatsApp “the leading surface where people engage with Meta AI” (Q2 2026 call); six months earlier Meta’s CFO had added that “in the U.S., Facebook is a stronger driver of engagement.” Since April 8, 2026 the assistant runs on Muse Spark, the first model from Meta Superintelligence Labs, which Meta says launched in the Meta AI app and on meta.ai before spreading to the other apps and, gradually, to Ray-Ban Meta and Oakley Meta glasses in the US and Canada.

What a source looks like depends on the door, and Meta documents only one pattern. Its help center says Meta AI “may use sources and links from the internet to inform its responses” and that you see them “by selecting Sources under its response.” Everything else in this table is what Meta has said about each surface — and, as often, what it hasn’t:

Two things follow. First, the answer arrives as a message — read on a phone, often in a group, sometimes forwarded. The brand named in the sentence is what everyone in the thread sees; the link sits one tap away in a sheet most people never open. Second, private chats still search. Incognito chats, which Meta launched on WhatsApp and the Meta AI app in Q2 2026 as conversations “even Meta can’t see,” start with web search on: the help center says it “begins toggled on” and sends “the search query to select search engines without connecting it to your account.” Part of your audience is conversations nobody can observe, Meta included.

02Two indexes: the open web and Meta's own apps

Meta AI doesn’t search every time. Meta’s search grounding docs, written for developers using the same models, say the model “skips the search when it can answer confidently from its training data,” and that simple factual questions “typically do not trigger a search.” When it does search, it has more than one place to look. On Muse Spark’s launch day, Meta AI listed its tools to Simon Willison, and a Meta engineer confirmed they belong to the new harness. Read in order, they’re the path of a searched answer:

1. 01

   browser.search writes a main query and alternatives

   The tool takes a primary_query, a list of alternative_queries, an optional since date, and verticals for news, sports, weather, finance, datetime and local. Willison’s write-up lists the engine behind it as undisclosed.

   Your move: Title pages and open them with the words people actually ask in: product names, places, and the year where freshness matters.

2. 02

   meta_1p.content_search searches posts on Meta's apps

   “Semantic search across Instagram, Threads, and Facebook posts,” built from “captions, visual analysis, transcripts,” over public posts and private ones the asker can see, from January 1, 2025. It filters by author, platform, place and date.

   Your move: Name your brand and products in captions, on screen and out loud: those are the three things the index is built from.

3. 03

   meta_1p.meta_catalog_search looks up products

   Shopping questions can “search for products in Meta’s product catalog,” and the app’s shopping mode searches Facebook Marketplace “alongside options from across the internet.”

   Your move: If you sell on Facebook or Instagram, keep your catalog’s names, prices and stock consistent with your site.

4. 04

   browser.open and browser.find read the page

   browser.open loads a result; browser.find looks for exact matches of a pattern inside it. Meta doesn’t document which user agent does the reading or whether it runs JavaScript.

   Your move: Put the answer in the raw HTML, early, in plain words an exact-match search can land on.

5. 05

   The answer cites a subset of what it read

   In Meta’s Model API the retrieved results are “every source the model considered” and the citations “the subset it actually cited.” In the apps, sources sit under the response.

   Your move: State one fact per sentence, so a citation has something whole to attach to.

The web side: partners Meta stopped naming

Microsoft announced in September 2023 that Bing would give Meta AI “more timely and up-to-date answers with access to real-time search information,” and Ray-Ban Meta’s release notes described glasses answers in January 2024 as “powered in part by Bing.” In April 2024 Google joined, though Search Engine Journal “could only get Meta AI to search using Bing” in its tests. Since then Meta has stopped naming names. Its AI terms, effective May 13, 2026, say Meta “may share certain information with select partners, like search engines,” and that in private chats “search terms based on your chat may be sent to these search partners.” You can’t optimize for a partner you can’t see, so being indexed and ranking well in both Google and Bing remains the broadest way into the candidate pool.

Meta's own crawler is building an index

The Information reported in October 2024 that Meta was “working on a search engine that crawls the web to provide conversational answers” (Social Media Today’s summary). meta-webindexer is the documented piece of it. In Promptwatch’s crawler logs it sat around 2% of AI crawler requests until July 17, 2026, then climbed to 37.8% on August 9 — about 17 times higher in under a month. On August 6 Pieter Levels posted that Meta staff had told him Meta is building its own index so its AI’s searches stay off Google; Search Engine Roundtable carried no comment from Meta. Treat the motive as unconfirmed. The crawl is real, and it’s the one Meta ties to citations.

Licensed news is linked by contract

Meta’s first AI news deal, with Reuters in October 2024, came with each such answer “expected to include a link to the Reuters story on which it’s based” (SiliconANGLE). In December 2025 Meta added CNN, Fox News, Fox Sports, Le Monde Group, People Inc., The Daily Caller, The Washington Examiner and USA Today, whose answers “will also include links to articles” (TechCrunch). For a news publisher, a deal is the documented route to linked answers; for everyone else, those outlets are the competition on current-events questions.

03Five Meta user agents, and the one tied to citations

Meta’s crawler page lists five agents, each sending a lowercase token such as meta-webindexer/1.1. It prefers “industry-standard practices like robots.txt” over “non-standard formats like NoAI tags,” and warns that its crawlers “may cache the contents of robots.txt for up to 24 hours.” There’s no FacebookBot on the current page, so an old rule naming it no longer maps to a documented Meta crawler. Flip each agent below to see the robots.txt group it needs and what you trade away:

# Meta AI search
User-agent: meta-webindexer
Allow: /
 
# User-requested fetches (may bypass robots.txt)
User-agent: meta-externalfetcher
Allow: /
 
# Model training
User-agent: meta-externalagent
Disallow: /
 
# Outside Meta AI (may bypass robots.txt)
User-agent: facebookexternalhit
Allow: /
 
# Outside Meta AI
User-agent: meta-externalads
Disallow: /

Two mechanics catch people out. A named group replaces your User-agent: * group for that agent rather than adding to it, so repeat any paths you still want blocked. And robots.txt is per host: shop.yoursite.com needs its own file.

Meta’s crawlers are heavy, which makes them tempting to throttle. Fastly attributed 52% of the AI crawler traffic it saw from mid-April to mid-July 2025 to Meta, against 23% for Google and 20% for OpenAI. Rate-limit if you must, but a 429 to meta-webindexer is a page Meta AI can’t cite.

To catch impostors, Meta’s page calls allowlisting by IP address “more secure” than by user agent, but links to no IP list. Meta’s network is AS32934. Meta doesn’t promise every crawler request comes from it, but it’s the first check when a meta-* user agent looks spoofed:

# Routes announced by Meta's network (AS32934)
whois -h whois.radb.net -- '-i origin AS32934' | grep ^route | head
 
# Who owns the address behind a suspicious hit?
whois 157.240.0.1 | grep -iE 'orgname|origin'

04What Meta's fetchers need from a page

Meta publishes no rendering, freshness or structured-data guidance for Meta AI. What’s documented is crawler access, the link-preview rules for facebookexternalhit, and how the search tool behaves in the API. The rest of this list is inference, marked as such.

meta-webindexer allowedRequired

In robots.txt and at the WAF. It’s the one agent Meta ties to citing and linking your content in Meta AI.

The answer in raw HTMLRequired

Undocumented for meta-webindexer, but Vercel saw Meta-ExternalAgent fetch pages without running JavaScript. Treat client-rendered text as invisible — see server-side rendering.

meta-externalfetcher not WAF-blockedHelps

It fetches at a user’s request and may bypass robots.txt anyway. A WAF block only means the answer gets built from someone else’s page.

The answer near the topHelps

browser.find looks for exact matches inside an opened page. A plain sentence that repeats the question’s key words, early on, is the easiest thing to land on.

Fast, error-free responsesHelps

A 429, 5xx or timeout to a Meta agent drops that page from the answer it was fetched for.

Open Graph in the first 1 MBHelps

Documented for facebookexternalhit: Open Graph tags “before the first 1 MB,” gzip and deflate support, and a page that can be crawled “within a few seconds.” It governs your card whenever the URL is shared on Facebook, Instagram or Messenger.

Structured dataUnconfirmed

No statement from Meta. Its tools hand the model page text, so keep every fact in visible copy, not only in JSON-LD.

Linked social profilesUnconfirmed

Post search filters by author. sameAs links from your Organization schema to your Facebook, Instagram and Threads accounts can’t hurt; Meta doesn’t document entity matching.

SitemapsUnconfirmed

Not documented for meta-webindexer. Keep one with honest lastmod values, referenced from robots.txt.

IndexNowNo effect

Meta isn’t on IndexNow’s list of participating engines. It still reaches Bing, a past Meta AI partner.

NoAI meta tagsNo effect

Meta says it uses robots.txt “rather than non-standard formats like NoAI tags.”

Test the two agents that matter with the user-agent strings Meta documents. The second command is the link-preview check from Meta’s own page:

# Does Meta's search crawler get your real content?
curl -s -A "meta-webindexer/1.1 (+/documentation/sharing/webmasters/web-crawlers)" \
  https://yoursite.com/pricing | grep -c "Plans start at"
# 0 = the text is rendered by JavaScript, or a WAF blocked the request
 
# Meta's link-preview simulation
curl -v --compressed -H "Range: bytes=0-524288" -H "Connection: close" \
  -A "facebookexternalhit/1.1 (+http://www.facebook.com/externalhit_uatext.php)" "$URL"

05Getting into the post index: captions, transcripts, Groups

Most assistants can only find you on the open web. Meta AI can also find you in posts on its own apps — Grok does the same with X, but Meta’s index spans three apps. At the Muse Spark launch Meta promised “Reels, photos, and posts woven directly into your answers, with credit back to the content creators,” and local answers showing “public posts from locals who know the area.” In June 2026 Facebook added AI Mode, whose answers are “grounded in what people are saying publicly across our apps like in Groups and Reels.” And on its Q4 2025 call Meta said it is “identifying the most relevant content across our platform to pull into responses.” Your posts, and other people’s posts about you, are retrieval candidates. Here’s what the tool definition and Meta’s announcements say about getting picked up:

06Measuring Meta AI without a console or a referrer

Meta publishes no webmaster console for Meta AI and doesn’t document the referrer its answer links carry. GA4’s AI Assistant channel covers “sources like ChatGPT, Gemini, Deepseek, Copilot, or Grok,” per Google’s channel documentation — not Meta AI. So measurement is three partial views stitched together: analytics, logs and a prompt panel. Here’s what to expect in analytics, by where the click starts. All of it is our read:

Give Meta AI a custom channel, placed above Referral so it claims sessions first. The first rule is clean; the second will over-count, because most link-shim traffic is ordinary Facebook sharing, so report it as a ceiling, not a number:

# Session source: Meta AI website
(^|\.)meta\.ai$
 
# Session source: Facebook link shims (Meta AI mixed with social)
^(l|lm|m)\.facebook\.com$

Your server logs are more honest. meta-webindexer hits show what Meta is indexing; meta-externalfetcher hits are pages fetched because a person asked about them — the closest thing to a live-demand signal Meta gives you. The AI crawler log analyzer does this in the browser; on a server:

# Meta agent hits, by agent
grep -oiE "meta-webindexer|meta-externalfetcher|meta-externalagent|meta-externalads|facebookexternalhit" access.log \
  | tr 'A-Z' 'a-z' | sort | uniq -c
 
# The pages people asked Meta AI to fetch
grep -i "meta-externalfetcher" access.log | awk '{print $7}' | sort | uniq -c | sort -rn | head -20

Then test retrieval directly. Meta’s Model API has a web_search tool, and its docs show how to return every result the model considered alongside the subset it cited. Run your target questions through it on a schedule — knowing the apps’ prompts, personalization and partners may differ from the API’s:

import os
from openai import OpenAI
 
client = OpenAI(base_url="https://api.meta.ai/v1", api_key=os.environ["MODEL_API_KEY"])
 
r = client.responses.create(
    model="muse-spark-1.3",
    input="What's a good carry-on backpack for a week in Lisbon?",
    tools=[{"type": "web_search"}],
    include=["web_search_call.results"],
)
 
# web_search_call.results = every URL retrieved
# url_citation annotations on output_text = the URLs cited

A manual panel, across surfaces

The API can’t show you the WhatsApp answer, so check a short list of prompts by hand once a month, in the same account, in WhatsApp and the Meta AI app, and note who is named and which sources appear:

1. Your brand as a question: “What is [brand] and is it any good?”
2. Your category with a constraint: “Best [category] for [use case] under [price]”
3. A comparison: “[brand] vs [competitor] for [use case]”
4. A local or situational ask, if it fits: “Where can I get [product] in [city]?”
5. A shopping ask in shopping mode, with and without an @ mention of your account

07What Meta hasn't said, and how to hedge each gap

Meta documents less about how its assistant retrieves and cites than OpenAI, Google or Perplexity do. These are the gaps that change what you’d do, and the move that holds up whichever way each one resolves:

One more agent to plan for: Meta launched Muse, a personal agent people can talk to in its own app or in WhatsApp, in the US on September 8, 2026, and it also owns Manus, acquired in December 2025. Neither is Meta AI, but both browse for people — and the pages that serve an agent well are the same answer-first, server-rendered pages this guide asks for.

The Meta AI to-do list

Everything above, in the order to do it. Ticks are saved in this browser, so you can come back to it.