New methodology Method 1.2.1 · 17 Nov 2026
A more accurate index: what changed, and why scores moved.
The Agentic Maturity Index now measures what an agent can actually get done through an interface, not how a vendor packages its interfaces. Four changes, each tested on frozen data before anything was published.
- Approval of each action is no longer a weakness. If the interface executes the change once a person approves it, the capability counts in full.
- A draft is not a change. An interface that only saves a draft for a person to publish elsewhere no longer counts for the buying tasks.
- Reach is described, not scored. Splitting one capability across several interfaces no longer earns extra points.
- Access comes from the interface that does the work. Not from a different, easier-to-reach one.
The level is lower because a whole component left the score, not because the market went backwards. Compare like with like: the series was restated under the new method, and the method 1.1 editions are kept unchanged in the archive.
Why we changed it
Two things the old method got wrong
After the index launched, vendors and practitioners pointed at two problems, and both held up when we checked them against the data.
It punished governance. Method 1.1 marked a capability "limited" whenever a person had to approve the action, which halved its points. But an agent that changes a budget after a person signs off has changed the budget. Approval is a control, and controls are what buyers want to see. Scoring them as missing capability rewarded the vendors with the fewest guardrails.
It rewarded packaging. Reach gave up to 20 points for having three or more agent interfaces. A company that splits one job across three servers scored higher than one that does the same job through one. That measured how a product is cut up, not what an agent can do with it.
The four changes
Before, now, and why it is more accurate
1. Approval is not reduced capability
- Before
- Any approval step made a capability "limited": half points.
- Now
- If the interface itself executes the documented change after a person approves each action, it counts as documented and is shown as "documented, needs approval". A standing switch that turns a class of actions on is pre-authorisation, not approval of each action, so it stays limited.
- Why
- The score now separates what the agent can do from how safely it is governed. Governance gets its own display instead of a penalty.
2. A draft is not an applied change
- Before
- An interface that saved draft campaigns or proposed budgets still qualified for "change a campaign" or "change bids and budgets", with a condition.
- Now
- It no longer qualifies. A person publishing the draft elsewhere is the change, not the agent.
- Why
- Task authority now means the agent's action reaches the live object. 3 draft-only capabilities were identified; none of them changes a score today, so this rule protects the index going forward rather than moving anyone now.
3. Reach is described, not scored
- Before
- 12, 16 or 20 points for one, two, or three or more agent interfaces.
- Now
- No points. The other four components keep their rules and scale up in proportion: capability depth 50, task authority 31.25, access 12.5, evidence 6.25, computed without intermediate rounding.
- Why
- Who an agent can reach matters, but counting servers did not measure it. Reach is now a separate view in real units: the ad platforms, publishers and operations a connection reaches, marked self-declared or verified, and "not published" where the vendor does not say.
4. Access comes from the interface that does the work
- Before
- The most open access tier of any of a company's agent interfaces, even one that earned no other points.
- Now
- The best access tier among the agent interfaces that earn capability or task points. If their access is not established, it earns nothing and says so.
- Why
- An open documentation server should not make a gated buying agent look easy to get. 7 companies change on this rule.
What moved
Who moved, and why
Both methods run here on the same data (2026-11-18.40), so every move below is the method, not new research. Most rank movement comes from removing reach: companies that spread their agent features across several interfaces fall, and focused single-interface companies rise.
Up the most
- Adspirer ▲14 rank 21 → 7 · 66 → 80a single agent interface, so removing reach costs it least; Adspirer Retail Media Ads write and Adspirer Retail Media Ads optimize now count as documented, with approval of each action.
- Zeotap ▲12 rank 53 → 41 · 48 → 51.3a single agent interface, so removing reach costs it least; Zeotap Agentic CDP MCP server activate now counts as documented, with approval of each action.
- Airbridge, AudienceMix, Google Cloud, Hightouch, Kochava, Lifesight, Roblox, Unity, Viomba ▲12 rank 71 → 59 · 43 → 38.8 eacha single agent interface, so removing reach costs it least.
- Mediahuis ▲11 rank 69 → 58 · 44 → 40a single agent interface, so removing reach costs it least.
- OneSignal, Sovrn ▲10 rank 64 → 54 · 46 → 42.5 eacha single agent interface, so removing reach costs it least.
- Nano Interactive, Roku ▲10 rank 66 → 56 · 45 → 41.3 eacha single agent interface, so removing reach costs it least.
Down the most
- iHeartMedia ▼38 rank 40 → 78 · 56 → 35reach no longer adds 20 points for 3 agent interfaces; access is not established on the interface that does the work.
- commercetools ▼36 rank 58 → 94 · 47 → 28.8reach no longer adds 16 points for 2 agent interfaces; access is not established on the interface that does the work.
- Google Ads ▼24 rank 52 → 76 · 49 → 36.3reach no longer adds 16 points for 2 agent interfaces; access now comes from Google Ads MCP server, the interface that does the work.
- Precisely ▼23 rank 69 → 92 · 44 → 30reach no longer adds 20 points for 3 agent interfaces.
- Nielsen ▼19 rank 82 → 101 · 41 → 26.3reach no longer adds 20 points for 3 agent interfaces.
- Nexxen ▼17 rank 119 → 136 · 32 → 19.4reach no longer adds 16 points for 2 agent interfaces.
Approval now counts: Adapty (optimize), Skai (activate), Adspirer (write, optimize), Zeotap (activate).
Access now follows the working interface: FreeWheel, Google Ads, iHeartMedia, commercetools, Adsquare, Horizon Media, Onetag.
Every company's Why panel on the index names the interface behind each point.
New, shown beside the score
Controls, test results and reach: shown, never scored
- Controls. For each agent interface, the approval policy and where it is enforced, tied to the specific operation reviewed. 60 interfaces document an approval policy for their main write or activation. "Not documented" never means "no controls exist".
- Test results. AgenticAdvertising.org's protocol conformance results, per track and AdCP version, with environment and date, for the 9 agent interfaces that map cleanly to its public registry. A result older than 30 days is shown as stale, never as passing. Conformance is not evidence of safe spend or delivery.
- Agent reach. For buyer agents, seller agents and gateways: who one connection reaches, in separate units, with its source.
These stay out of the score until there is a way to test them the same way for every vendor: for example, whether a denied or expired approval actually stops the action.
How we made the change
Locked before the results
- Froze the data and the list of scored companies first.
- Wrote the new rules, weights and access formula down and committed them before computing a single new score.
- Labelled every one of the 210 limited agent capabilities twice, in two independent AI-assisted passes over the Atlas record and its cited sources. A capability changed only when both passes agreed. The 28 where they disagreed or the record does not say stay as they were.
- Measured each change on its own, then together, and checked the ranking against ±20% changes to every weight: the top ten did not change.
- Made one amendment after the review and before publishing: approval means approval of each action, so a standing opt-in switch stays limited. It is recorded as version 1.2.1 rather than edited into 1.2.
- Kept a same-day data correction apart: OpenX's Prebid floor no longer counts as a bids or budgets control. It does not affect the index.
The rubric, the comparison, every label and both passes are public in the method files. Published editions are never edited: the method 1.1 editions sit unchanged in the history archive, and the new series is restated from the Atlas's own history and labelled as a reconstruction.
Still true
What the index does not measure
It measures documented agent capability and authority. Not product maturity, safety or quality, and not production performance. A company's score covers capabilities across its interfaces; it does not mean one connection can do the whole workflow. Missing evidence still earns nothing, and that is not a finding that the capability is absent.