measurement.some of your visits aren't people anymore. separate them before reading conversion.
Muse, Meta's personal agent, visits business websites on behalf of its users, and Meta is already proposing a standard for it to do so with permission. For anyone measuring a site, the question is shorter: how many of those visits are people.
in three lines
- A report by HUMAN, a company that sells agent detection, attributes to Muse 72% of the daily AI agent requests it observes on websites. It is a vendor figure, unaudited.
- On 6 October Meta and Sierra introduced the Personal Agent Protocol, a draft standard for personal agents to identify themselves and request permissions on business websites.
- Until that standard exists, an agent visit can land in the same metrics as a human one. Before reading a conversion rate, you need to know how much of the traffic isn't people.
what happened and when
On 8 October PPC Land analysed a report HUMAN published on 28 September, with data through the 27th. According to the company, Muse accounts on average for 72% of the daily AI agent requests its platform records on websites, far ahead of the next agent.
The number needs careful reading. It comes from a company that sells agent detection, it is drawn from its own customers and, according to PPC Land, the report doesn't say how many requests it measures, over what period, or what counts as a request. It signals direction, not the size of the phenomenon.
an agent that walks into other people's sites
Muse isn't a search engine indexing pages: it browses on a person's behalf, checks, compares and sometimes tries to buy. That has already caused friction. According to coverage from those days, Amazon cut off its access to the store on 20 September, arguing among other things that the agent didn't identify itself while browsing.
Meta's answer came on 6 October: together with Sierra it introduced the Personal Agent Protocol, a standard for personal agents to authenticate on business websites and for the customer to decide whether they get read-only or write access. For now it is a draft: the first specification hasn't been published and payments are left out of the first version.
why it matters to whoever measures
An analytics tool counts sessions. If an agent opened a session and nobody tells it apart, it goes into the same denominator as people's visits: visits go up, the conversion rate goes down, and nobody changed anything on the site. HUMAN argues that treating those sessions as human distorts conversion, funnels and attribution. That is the thesis of the company selling the fix, but the mechanism is arithmetic.
It hits hardest where there is no online sale. When the conversion is a form, a download or a click to WhatsApp, volumes are small, and a few dozen visits that aren't people are enough to move the rate from one month to the next.
how we work with it
When we read a site, the first question isn't how much it converted but who visited it. A jump in visits without a jump in conversions, or with a falling rate, gets checked by source, by user agent and by behaviour —pages per session, time on site, hours— before it is explained as a problem with the page.
The conversions that matter are measured with events a person performs and that can be verified on the other end: a form that arrives, a message that gets answered. And when a report can't separate agent traffic, it says so: a number that mixes two things isn't presented as if it measured one.
An agent browsing a site counts as one more visit unless someone separates it. Before reading a conversion rate, you need to know how much of that traffic isn't people.
what changes in measurement
the conversion rate needs context
If visits go up and conversions don't, the first step is to check where that traffic comes from, not to redesign the page.
the conversion that counts is the verifiable one
A form received or a message answered can be checked on the other end. A visit to a thank-you page, not always.
vendor data gets cited as such
The 72% is a figure from a company that sells agent detection, with no independent audit. It is useful for watching the trend, not for sizing your own site.
blocking is a decision too
You can shut agents out, but it has a cost: if a person uses one to compare options, the site that blocks it is left out of the comparison. Decide it; don't leave it to the default anti-bot setting.
questions we get
Does Muse show up in my analytics as one more visit?
The coverage we read doesn't detail how Muse identifies itself or how analytics tools classify it. So it is better to check your own data by source and behaviour than to assume it is already filtered out.
Should we block agents?
Not necessarily. Blocking them keeps them from distorting measurement, but it also takes the site out of the tasks a person hands to their agent. It is a business decision, not a default technical setting.
Does the 72% apply to my site?
No. It is the split between agents that one vendor sees across its customers, not the share of agent traffic on any given site. Your site's share only comes from your own data.
where this comes from
- HUMAN report on Muse traffic, 28 September 2026, and its analysis in PPC Land, 8 October 2026.
- Introduction of the Personal Agent Protocol by Meta and Sierra, Meta for Business, 6 October 2026, and its coverage in Implicator, 6 October 2026.
- Amazon blocking Muse from its store, Implicator, 6 October 2026.
- Daily sweeps by AIRadar and SocialRadar, Champe Agency's own agents, 3 to 9 October 2026.
Sources are cited as text, with outlet and date. We don't link addresses we haven't verified.
and what it takes
The note explains the problem; the services explain the work.
the other notes
permissions for agents
Meta introduced its agent platform for businesses and opened Muse to them. Connecting an agent to an account is the easy part. The decision that matters is what it can do there and who finds out when it does.
read the note → 09 · September 26, 2026the facebook link quota
With Meta One, an external link on a Facebook Page went from being a low-reach format to being a rationed resource. The question is no longer how to add the link: it is which posts deserve one.
read the note → 08 · September 19, 2026the unranked feed
In a single week, three governments moved on the same front: who can hold an account, which feed they see and which ads reach them. For a brand the effect is neither moral nor legal: it is about reach and measurement.
read the note → 07 · September 12, 2026declaring ai use
Instagram flags barely retouched real photos as AI and lets generated images through. If detection gets it wrong in both directions, disclosure cannot depend on it.
read the note → 06 · August 28, 2026running on agents
Plenty gets said about agencies using AI, and very little about how one is actually run. This is the method from the inside: how knowledge is organised, where the human sits, and what breaks when nobody looks after it.
read the note → 05 · August 28, 2026reading year-over-year comparisons
Global CPM is up 32% since 2024 and US inventory is at an all-time high. A campaign can look worse in this year's report without having gotten worse for a single day.
read the note → 04 · August 28, 2026originality and reach
Instagram no longer rewards originality out of principle. It measures it. Reposting loses reach in ways you can verify, the penalty has moved beyond video, and the hashtag no longer makes up for anything. Here is what counts.
read the note → 03 · August 28, 2026measuring without e-commerce
Most real work doesn't end in a cart. It ends on a shelf, in a store, in a doctor's appointment. That can be measured. What can't be done is inventing the metric at month end.
read the note → 02 · August 28, 2026meta placement controls
Meta has finished retiring the manual placement selector. What remains are value rules with a hard floor of -90%. Brand safety stops being a list of unchecked boxes and becomes a production decision.
read the note → 01 · August 28, 2026creative volume on meta
Meta split the decision of which ad you see into two stages. The first one reads the creative, not the targeting. That is where creative volume stopped being a production luxury.
read the note →