Meta's Muse reads 25 sources and answers in one paragraph

Meta launched Muse in September: a personal agent with its own browser that researches products, compares them and, once the buyer approves, buys them. On 28 October Jason Saltzman, who leads research at AirOps, posted the first look at how it answers buying questions next to ChatGPT, Gemini and Google’s AI Mode, from answers collected between 23 and 27 October.

It’s one chart with six numbers. I spent a morning with them. One is bigger than it looks, two need a second look, and together they describe a part of the buying journey that analytics wasn’t built to see.

The answer is one paragraph

777 characters is one paragraph, about four times as long as this one. ChatGPT’s average answer to the same kind of question is almost seven times as long as Muse’s, and Gemini and AI Mode land in between.

Length alone wouldn’t matter much. What matters is what fits in it.

The other three answer a buying question the way a comparison site does: a list of options, often a table, sometimes a winner on top. Muse answers it the way a friend does, in a few sentences with a recommendation.

And it doesn’t stop to ask what the buyer needs.

ChatGPT ends almost three quarters of its answers with a question to the buyer, and AI Mode nearly all of them. Muse does it in 0.2%. AirOps doesn’t say whether the suggestion buttons Muse shows under some answers were counted; a reader asked under the post and got no answer. Even so, the gap is wide.

Put the three together and the shape is clear: a short answer, no list, no question back. Muse decides.

On a list, the brand in third place still gets read. In one paragraph, third place doesn’t exist.

It reads more than any of them

Here’s the other side of the short answer. Muse cites 25.1 sources per answer, 3.4 to 4 times as many as the other three. And 17.6% of them are review sites and user forums, against 3.2% to 9.1% for the others.

Multiply the two and the gap gets bigger: about 4.4 review or forum pages behind every Muse answer, against 0.2 for ChatGPT. That’s about 19 times as many. Jason added one more number in the comments: about a third of what Muse pulls comes from sites no other engine surfaced for the same prompts.

The review thread you haven't read may be the page that decided for your buyer.

So Muse reads a lot, much of it written by other buyers, and much of it on sites the other three didn’t surface. Then it compresses all of that into one paragraph.

Two numbers to read with care

The chart is a first look, and AirOps hasn’t published how it was made yet: how many questions, in which categories, through which Muse surface. Two of the numbers need a second look before anyone builds a plan on them.

“A much broader set of sources” is mostly more sources. Muse cited 10,171 different websites in the sample, against 2,372 to 2,660 for the others. That looks like four times the breadth. But the number of different sites always grows with the number of citations, and Muse cites 3.6 times as many per answer as ChatGPT. Divide one by the other and most of the gap goes away.

Per source cited, Muse reaches 6% to 16% more different websites than the other three. It reads more, and about as widely as they do. The same arithmetic tells us one thing about the sample for free: to cite 10,171 different websites at 25.1 per answer, AirOps needed at least about 400 Muse answers.

25 sources in 777 characters. That’s one source for every 31 characters, three or four per sentence. No answer footnotes itself that densely. eesel, which tested Muse as a research tool, reports that it doesn’t cite sources by default and can list every page it visited instead. So the likeliest reading is that Muse’s 25.1 are pages it opened while researching, and ChatGPT’s 6.9 are sources it showed in the answer. Both are worth knowing, but they’re different things, and the comparison will mean more once AirOps says how each was counted.

One more caveat: Muse is an agent with its own browser, compared here with three answer engines. Asked about a comparison with the other agents, OpenAI’s Dots and Anthropic’s Claude Duet among them, Jason replied “stay tuned”.

None of this changes the shape. A short answer, no list, no question back, and a lot of reading behind it. That holds however the sources were counted.

Where the visit goes

Then Muse acts. It runs a full browser in a cloud machine, opens product pages, checks price and stock, and fills the cart for the buyer to approve. Here’s what that looks like from the brand’s side, from the companies that watch web traffic for a living:

• It doesn’t say who it is. Meta hasn’t published a user agent or an IP range for Muse. Its browser looks like desktop Chrome on Linux (Botify), and 99.3% of its requests in DataDome’s sample came from cloud and CDN networks.
• It’s already most of the agent traffic. In HUMAN Security’s early observations, Muse made 72% of daily AI agent requests, and 84% of them went to product and search pages. DataDome counted 29.1 million Muse requests in its first 12 days of full availability.
• It gets further than other agents. HUMAN found Muse 25% more likely than other AI agents to reach a checkout page.
• Amazon blocked it on 21 September, saying it didn’t authorise Muse to shop there (Adweek).

And here’s what it looks like in your analytics, from Botify’s write-up:

• Muse runs the page’s scripts, so the visit is recorded like a person’s, and with no referrer it usually lands in direct.
• When Muse fetches a page’s text without running scripts, the request reaches the server but not analytics.
• When the order goes through a connector or Shop Pay, it can skip the checkout pages altogether, so the revenue arrives with no session behind it. Shopify merchants get a Meta sales channel line; other merchants get revenue with no source.

If you run GA4, Botify’s signs to look for are more direct traffic, more desktop Chrome on Linux, a lower conversion rate on direct, deep product page sessions that buy within minutes, and more orders with no session.

The dark funnel just got literal

Marketers already have a name for the part of buying they can’t track: the dark funnel. The podcast someone heard, the private chat where a colleague recommended a vendor, the forum thread read on a phone. You see the result, a branded search or a direct visit, but not the reason.

Muse adds a new room to it, and the numbers above say how dark it is. Behind each answer the agent reads about 25 pages for the buyer, many of them reviews and forums, a third of them on sites the other assistants didn’t surface. It decides in one paragraph, without asking the buyer what matters to them. Then it visits your site looking like a person, or doesn’t visit at all.

Three things follow for a brand.

Your reports show the result and hide the reason. A rise in direct traffic, a dip in conversion rate, an order without a source. Even when you can tell an agent was there, nothing in your data says what it told the buyer about you, which competitor it put next to you, or why it chose them.

Your analytics can show that the agent came. It can’t show what the agent said about you before it came.

Second place earns nothing. On a list, the brand in third place still gets read and sometimes clicked. In one paragraph with one recommendation, the brand the agent picks gets the visit and the sale, and every other brand gets nothing, not even a lost visit to count. The “visibility” dashboard can stay green the whole time while the bottom line is getting deep in the red.

The pages that decide often aren’t yours. Review sites and forums carry about 4.4 of Muse’s sources per answer, and a third of what it reads came from sites the other three didn’t surface. What works for a brand in ChatGPT may do nothing in Muse, and the other way round.

The funnel is dark to your analytics. The AI assistants themselves can still be asked. What they say about your brand when a buyer researches, compares and decides, which competitor they pick instead, and the reasons they give: that’s where the sale is won or lost now, and it can be measured today.

By the time it shows up in revenue, it’s already set in the answers.

We do this research for brands. If you want it for yours, start with the free assessment.

Sources: AirOps Research, "Muse is 'built different'", Muse against ChatGPT, Gemini and AI Mode, 23 to 27 October 2026, published by Jason Saltzman on LinkedIn, 28 October 2026, and his replies under the post; the six figures are read from the published chart. Botify, Jordan Long, "Meta Muse", 21 October 2026. HUMAN Security, early observations of Muse, 28 September 2026, as reported by PPC Land, 8 October 2026. DataDome, AI Traffic Report Q3 2026, as reported by Help Net Security, 19 October 2026. Adweek, "Amazon locks out Meta's Muse in agentic shopping standoff", 21 September 2026. eesel, "Meta Muse for research", 8 October 2026. Derived figures (sources per answer by type, the per source index, the minimum sample) are our arithmetic from AirOps' published values.