--- title: "Agent Success Score: measuring AI visibility and usability | Hyperize" description: "How Hyperize measures whether AI agents find an offer and complete a task: visibility, usability, evidence and a score for each agent role." canonical: https://www.hyperize.ai/en/methodology/agent-success-score lang: en last-updated: 2026-10-04 --- # Agent Success Score. Methodology · Concept The Agent Success Score measures how well a brand enables AI agents to complete a defined customer task. It combines AI Visibility, AI Usability and Evidence into a value from 0 to 10, separately for each agent role. The underlying tests show what worked, where agents stopped and under which conditions. The score summarizes the test. The evidence explains the result. What it is not A brand metric, not a support metric. We assess the offer an incoming AI agent uses: a website, its data and permitted functions. A contact-center agent success rate instead measures how often support agents resolve an issue. This is neither a general benchmark of AI models nor a success rate for a support team. Why this exists ## Measure the customer journey. A valid file or correct markup shows that a technical requirement is met. To understand the customer journey, we also test whether agents find the offer, read the relevant facts and reach the agreed endpoint \[S1\]. The task and test limits are fixed in advance. A technical barrier, a missing permission and a deliberate stop before submission are recorded separately. A failed run alone does not establish its cause or prove a lost sale \[S2\]. How it is built ## Two gates. Both must open. Being recommended does not mean an agent can use an offer. A usable offer may still go undiscovered. We measure both questions separately, with Evidence as the supporting layer. AI Visibility · 20% ### Does AI find and recommend your offer? We test task-related questions on the documented AI platforms. Unbranded questions show whether a brand appears on its own. Brand-seeded queries must be reported separately. AI Usability · 70% ### Can the agent proceed on your website? We test which agent classes can read facts and perform permitted steps. The value is derived from the observed access profile and the endpoint reached. It does not replace the test description \[S1\]. Evidence · 10% ### Can the result be traced to evidence? Evidence records what supports the finding and how consistent the tested results are. It complements the two measurement axes. A high component value does not remove a limited test scope. Composition · published Agent Success Score = (AI Visibility × 0.20) + (AI Usability × 0.70) + (Evidence × 0.10) Each input uses a 0–100 scale; the weighted sum is divided by ten. Published DHL example from Q2 2026: Visibility 46.04, Usability 68 and Evidence 85 yield 6.5 of 10 after rounding. One parcel task, confidence C; checkout behind login was not tested. This is a historical example, not a new measurement \[S3\]. Why trust it Two complementary checks, one finding. A query audit tests discovery. Agent runs test access and actions on the live offer. These methods examine different parts of the journey; agreement can strengthen a finding but does not remove their limits. Tasks are fixed before measurement, and published results must be read with their scope and date \[S2\]. The lane model ## One score per lane. Never blended. A shopping agent has a different job from a recruiting or press agent. The model distinguishes six roles, called lanes. A result for one role cannot stand in for the others \[S4\]. Commerce Buy Talent Apply After-sales Resolve Procurement Source Investor Research Press Research Each fully measured role has its own score. Missing measurements stay explicit. The DAX 40 comparison uses the shared Commerce test set; separate Talent tests are also published. Those application tests alone do not establish a complete Talent score \[S4\]. On scope The composition, not the derivation. The definition and weights are public. Detailed scoring logic and raw run records remain within the engagement. Results apply to the tested task, access route, systems and date. A 6.5 score is not a 65% probability of purchase. Seven self-test prompts provide observations, not this score. Rankings, citations and future task completion are not guaranteed. [Discuss my customer task →](/en/contact) Related ## The cluster around this metric. [ Fair test tasks ### Task Selection. Which tasks allow a fair comparison and how we distinguish direct routes from intermediaries. Read the rules](/en/methodology/task-selection) [ Technical implementation ### Your offer for AI. Pages, data and functions agents can read and use. Turn a finding into the right implementation. Understand my integration](/en/methodology/agent-surface) [ Public examples ### DAX 40 Agent Success Index. The methodology applied to real brands: measured tasks, results and limits in context. See the results](/en/dax40-index) Questions ## What people ask. What is the Agent Success Score? The Agent Success Score measures how well a brand enables AI agents to complete a defined customer task. It combines AI Visibility, AI Usability and Evidence into a value from 0 to 10, separately for each agent role. The underlying tests show what worked, where agents stopped and under which conditions. Is it a service bot's success rate? We assess the offer an incoming AI agent uses: a website, its data and permitted functions. A contact-center agent success rate instead measures how often support agents resolve an issue. How is the Agent Success Score calculated? AI Visibility contributes 20 percent, AI Usability 70 percent and Evidence 10 percent. Each input ranges from 0 to 100. The weighted sum is divided by ten and reported for each agent role. Usability is derived from the observed access profile; detailed scoring logic remains within the engagement. Sources ## Evidence and provenance. S1 Internal methodology Internal scoring definition Hyperize · May 2026 Internal definition, not publicly accessible. Supports: The two measurement axes, derived usability and public composition weights. Not an independently accessible source. S2 Methodology Task Selection Doctrine Hyperize · May 2026 [https://www.hyperize.ai/en/methodology/task-selection](https://www.hyperize.ai/en/methodology/task-selection) Supports: Rules for selecting fair tasks and separating direct routes from intermediaries. S3 Historical measurement DHL Group: published test Hyperize · Q2 2026 [https://www.hyperize.ai/en/dax40-index/brands/dhl](https://www.hyperize.ai/en/dax40-index/brands/dhl) Supports: Each input uses a 0–100 scale; the weighted sum is divided by ten. Published DHL example from Q2 2026: Visibility 46.04, Usability 68 and Evidence 85 yield 6.5 of 10 after rounding. One parcel task, confidence C; checkout behind login was not tested. This is a historical example, not a new measurement \[S3\]. S4 Measurement overview DAX 40 Agent Success Index Hyperize · Published measurement date per test [https://www.hyperize.ai/en/dax40-index](https://www.hyperize.ai/en/dax40-index) Supports: The shared Commerce test set and separately reported application tests. This explanation does not remeasure any brand. The internal scoring definition is not a public source. Linked pages show the public task-selection rules and historical examples. This page updates the explanation; it does not remeasure any brand. Last updated 2026-10-04 Next review 2026-11-03 Evidence tier proprietary Confidence A --- Canonical: https://www.hyperize.ai/en/methodology/agent-success-score Machine-readable index: https://www.hyperize.ai/llms.txt