Why I compiled this Agentic Search Optimisation Glossary
This is the companion to my AI Search Optimisation Glossary. That one covers being found and cited and this one covers what happens when the thing arriving at your site is an agent trying to finish a job. I have spent this year testing the agentic protocols on live sites and the shift I keep running into is people handing work over: Asking an assistant to compare, configure, book or buy and increasingly pointing it at a tool or another agent to do the actual doing. The data backs that up.
- HUMAN Security measured agentic traffic growing 7,851% year on year across 2025, from more than a quadrillion digital interactions.
- Cloudflare put bots at more than half of all internet traffic, though that counts every crawler and scraper of any kind, so agents are only a slice of it.
- Adobe found 39% of US consumers had used AI for online shopping, in a survey of more than 5,000 people.
- In HUMAN’s June 2026 census, 79% of agent activity landed on product and search pages and 2.34% reached checkout.
- Accenture asked 25,590 consumers: 74% would delegate a routine task and 9% would allow a fully autonomous purchase.
So agents are researching at scale and buying at the margins, which is exactly the window in which to get a site ready cheaply. There are 104 terms here across 6 ranges. Every standard and product carries a status flag in bold at the front of its entry, so you can see at a glance what is in production, what is still a proposal and what has already been quietly shut down. The first range is enormous and the rest are not, because about a fifth of these terms begin with the word agent and nearly half contain it somewhere.
A to B
A2A (Agent2Agent Protocol)
Production. Agent2Agent is an open protocol for one agent to negotiate with another, so a customer’s agent can talk to a brand’s agent directly a website. In practice it means publishing an agent card and an endpoint. It reached v1.0 with 150+ supporting organisations in its first year, which makes it the most credible of the agent-to-agent standards. So far it is an enterprise back-office story and the production deployments are in supply chain, financial services and IT ops.
Accessibility tree
The semantic representation of a page that browsers already expose for screen readers. Browser agents increasingly read it directly, which saves them interpreting pixels, which quietly turns years of accessibility work into agent-readiness work. The most satisfying overlap in this whole field and the one nobody planned.
ACP (Agentic Commerce Protocol)
Beta. OpenAI and Stripe’s open specification for agents to discover products and complete purchases with a merchant, covering cart, feed, orders and authentication. The spec is still in beta and still maintained by its two founders, with a neutral foundation described as an intention rather than a fact. It survived the Instant Checkout pullback, which is the more interesting detail.
Agent
A system that plans and carries out multi-step tasks using tools, going beyond producing text. The practical shift for marketers: An agent can compare, configure, book and buy, so your website now has to support finishing a job as well as describing one.
Agent allowlisting
Deciding which agents may act on your site, which is a different decision from which crawlers may read it and it carries commercial consequences, because blocking a shopping agent blocks a buyer.
Agent card
The JSON document an A2A agent publishes to describe what it is, what it can do and how to reach it, which works as a machine-readable capability statement.
Agent evals
Repeatable tests scoring whether an agent finishes a task correctly. Vendors quote them constantly, so ask what the task set was and who wrote it. An eval built by the people selling the agent tells you less than you would like.
Agent experience (AX)
The emerging name for designing a site so an agent can complete tasks on it, sitting alongside UX. I am wary of the acronym, though the underlying work is real, because most of it is ordinary clarity: Stable selectors, honest labels, forms that survive without hover states.
Agent gateway
Production. Infrastructure sitting between agents and your systems, handling authentication, rate limits and routing in one place. One of the four projects the Agentic AI Foundation holds and something your platform team will own.
Agent identity
The problem of proving which agent is acting and on whose behalf. It splits into two: Verifying the agent itself, which Web Bot Auth addresses and verifying the human authority behind it, which mandates and verifiable credentials address.
Agent memory
Persistent context an assistant carries between sessions about a user’s preferences, past purchases and constraints. It creates a personalisation layer no prompt-set test can see, because your tracked prompts have no history and your customers do.
Agent readiness
Whether an agent can actually complete a real, commercially valuable task on your site. The honest way to find out is to ask one to do it and watch where it fails. Most failures turn out to be ordinary website failures, which is why I would fix those before reaching for any protocol.
Agent session
An interaction driven by an agent. Most analytics stacks cannot separate one from a human visit, which is the root of nearly every measurement complaint in this glossary.
Agent share of traffic
The proportion of your visits that arrive from agents. Cloudflare reported more than half of internet traffic is now non-human and that as of June 2026 52% of all crawler requests are for AI training, with mixed-use crawlers a further 36%. Your own share is knowable from logs and from nowhere else.
Agentic AI Foundation (AAIF)
Production. The Linux Foundation body formed in December 2025 that now holds MCP, goose, AGENTS.md and agentgateway under neutral governance, with AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft and OpenAI among its platinum members. It is the answer to the “who owns this standard” question a stakeholder will eventually ask. A2A‘s proposal to join was still awaiting a governing board vote as of August 2026.
Agentic checkout
A purchase completed by an agent inside an assistant, without the customer visiting your checkout. The commercial question it raises is whose customer that is and nobody has answered it yet.
Agentic commerce
Agents completing end-to-end commercial journeys, from discovery through checkout and sometimes post-purchase support. Adobe measured AI traffic to US retail sites up 393% year on year in Q1 2026 and more usefully, that traffic converting 42% better than non-AI sources in March 2026 having converted 38% worse a year earlier. The visit went from roughly 0.62 times non-AI conversion to 1.42 times it in 12 months. Worth stating plainly that this is AI referral traffic, meaning humans clicking through from an assistant and not agents acting on a site.
Agentic loop
The cycle an agent runs: Plan, call a tool, read the result, decide what to do next, repeat until it finishes or gets stuck. It is why an agent can sail through browsing and searching and then fail at the basket, because each step depends on the one before it and a single broken link in that chain ends the task.
Agentic RAG
A retrieval loop where the model decides iteratively what to search next based on what it has already found. It is the mechanism behind deep research modes.
Agentic SEO
Making a site something an agent can act on: Clean product data, structured actions, forms that survive automation and protocols where they earn their place. It sits alongside the AI search acronyms as another branch of the same discipline.
Agent-originated conversion
A sale or signup where an agent did the work. Almost nobody can report it today, because the session looks human, so the honest answer to “how much revenue comes from agents” is that you cannot tell.
AGENTS.md
Production. A plain markdown file at the root of a code repository telling coding agents how to build, test and navigate it. Adopted by more than 60,000 open-source repositories and now held by the Agentic AI Foundation. It is the one file in this glossary with demonstrated adoption, which is worth sitting with given how much of the rest is still a proposal.
AI browser
Churning. A browser with a resident agent that can read, summarise and act on pages. The category is real and the products have churned: Perplexity’s Comet is live and free across desktop and mobile, Gemini in Chrome keeps expanding and OpenAI announced on 9 July 2026 that it would sunset the standalone Atlas browser, which stops working on 9 August 2026, with its capabilities folded into the ChatGPT desktop app and Chrome extension.
Amazon versus Perplexity
Live case. The case that decides whether you can block an agent. Amazon sued Perplexity over Comet shopping on its site and won a preliminary injunction, which the Ninth Circuit vacated in August 2026, reasoning that the Comet user was the one accessing Amazon’s computers because the browser requires user direction. The holding is narrow, on computer-access law and these specific facts and the underlying case continues. Its practical effect is to move the fight from access law to your terms of service.
AP2 (Agent Payments Protocol)
Production. Google’s protocol for establishing payment trust through cryptographically signed mandates proving a user authorised an agent’s purchase. It is at v0.2 with 60+ supporting organisations, positioned as an open extension to A2A and MCP, with standardisation running through the FIDO Alliance. It sits with the FIDO Alliance and not the Linux Foundation, which people get wrong.
ARD (Agentic Resource Discovery)
Proposal. Agentic Resource Discovery is a single JSON catalogue at a well-known path listing the resources and capabilities you deliberately expose to agents. Pilot it when agents cannot find things you have gone to the trouble of publishing for them. The spec was still a v0.9 draft as of May 2026.
Authentication for agents
How an agent proves it may access an account. It is the wall most agent tasks hit, because logins, two-factor prompts and CAPTCHAs were built to stop exactly this behaviour. Telling a customer’s agent apart from an attacker is genuinely hard.
Autonomy level
How far an agent may go before asking a person. It runs from suggesting, through acting with confirmation, to acting alone. Most production deployments sit at the middle setting, for good reason.
Bot versus agent
The distinction your infrastructure probably does not make. A bot crawls to build an index or a training set. An agent arrives because a specific person asked it to do a specific thing, which turns blocking it into a commercial decision.
Browser agent
An agent that drives a real browser to navigate, click, fill forms and complete tasks on live sites. Google’s AI optimisation guidance notes these can inspect rendered pages, the DOM and accessibility trees, which moves the question from whether a machine can read your page to whether it can finish a job on it.
C to E
CAPTCHA
The challenge designed to prove a visitor is human and now the most common hard stop in an agent journey. Decide deliberately where yours sit, because one on a product page is a very different commercial choice from one at password reset.
Chain of thought
The intermediate reasoning an agent produces before acting. In search surfaces it is what generates the sub-queries; in agents it is what generates the plan.
Chargeback liability
Who carries the loss when an agent buys the wrong thing. As of August 2026 no card network has published a settled rule for agent-initiated purchases. Raise it with your payments team before your merchandising team gets excited.
ChatGPT agent
Production. OpenAI’s agent mode, able to research and act using tools and a browser. It is the most widely available agent your customers actually have, which makes it the sensible one to run your first readiness test with.
Checkout agent
An agent that completes a purchase on the customer’s behalf. Its success depends entirely on whether your checkout is machine-navigable, which for most retailers it is not, because checkouts are the most aggressively bot-protected part of the site by design.
Claude for Chrome
Beta. Anthropic’s browser extension letting Claude act inside your browser. Still officially in beta, now open to all paid plans without a waitlist. Useful as a way to run an agent readiness test on your own site without building anything.
Cloaking risk for agents
Serving agents a different version of your page from the one humans see. Several vendors now sell this as machine-readable page variants. The idea is reasonable and the execution history is grim, because serving different content to machines has ended badly for a very long time. Keep the substance identical and change only the presentation.
Commission on agentic checkout
What a platform takes when an agent completes a sale inside its own surface. It is the first question a commercial director asks about Instant Checkout or UCP and the hardest to get a straight answer to, because the terms are still being set deal by deal.
Computer use
A model controlling a whole desktop through screenshots and synthetic input, so it can drive any application on the machine. Broader than a browser agent and correspondingly slower, more expensive and more failure-prone.
Context engineering
Deciding what goes into an agent’s limited context window and in what order. It has largely replaced prompt engineering as the term practitioners use, because with agents the hard part is managing what the model can see.
Copilot Actions
Preview. Microsoft’s agentic feature for Windows, still in preview through Copilot Labs and running inside a sandboxed agent workspace. Preview has lasted a while, so treat roadmap claims about it carefully.
Deep research
A mode where an assistant browses many sources over several minutes and returns a long report with an extended source list. It behaves like an agent and gets measured like a search surface, which is why it falls between these two glossaries and gets tracked by neither.
Delegated authority
The permission an agent carries from the person it acts for and the thing a merchant actually needs to verify before honouring an agent-initiated action. Mandates and verifiable credentials are the two mechanisms competing to express it.
Direct offers
Merchant-controlled product and pricing data supplied straight to an agentic surface, so the agent reads your numbers straight from you, with no scraping and no guessing. Where the commerce protocols are heading and the reason feed accuracy is about to outrank product copy.
DOM (Document Object Model)
The parsed, structured version of a page after the browser has built it. An agent reading the DOM sees whatever your JavaScript produced. Class names that change on every deploy and controls that only appear on hover, are the two things I see break agent tasks most often.
Elicitation
The MCP mechanism by which a server can pause and ask the user for more information mid-task. It is the protocol-level version of an agent admitting it is missing something and it is more useful than it sounds.
EntityMap
Stable spec, thin evidence. Two files at your domain root, entitymap.json plus an HTML companion, setting out your entities, the evidence behind them and how they relate. Stable at v1.0 since April 2026 with a validator, so further along than ARD or OKF and Chris Green puts it on the knowledge side. The only file here with a published result: Waikay’s 19-week test took weekly Copilot citations from 167 to 614, product pages up 406%, on its own site with no control group and Waikay built the standard. Dixon Jones has the how-to.
Form accessibility for agents
Whether an agent can complete your forms: Real labels, sensible autocomplete attributes, validation messages that say what is wrong and no dependence on hover or drag. It is the same list as human accessibility, which is why this is the cheapest agent-readiness work available.
Gemini Agent
Production. Google’s agent for complex multi-step tasks and where Project Mariner‘s technology ended up after that experiment closed.
Gemini in Chrome
Production. Google’s assistant built into the browser, expanding across markets through 2026 and being wired into a persistent browsing agent. Because it ships with Chrome, it is the most likely route to agentic capability arriving switched on by default, which is one of the signals I said would raise this work’s priority.
Google Business Agent
Production. Google’s branded shopping agent, letting shoppers chat with a retailer in that brand’s voice directly on Search. Eligible US retailers activate and customise it in Merchant Center, with Lowe’s, Michael’s, Poshmark and Reebok among the early users. It is a retail product, though the name leads plenty of people to assume it is a local-business one.
Google-Agent
Production. A user agent Google introduced for its AI agents browsing on a person’s behalf. As a user-triggered fetcher it does not obey robots.txt, on the logic that a person asked for it. Check whether your bot rules treat it as a crawler, because it is not one.
goose
Production. An open-source agent framework that started at Block and is now one of the four projects held by the Agentic AI Foundation. Included because you will meet the name in governance discussions and wonder what it is.
Guardrails
The filters and policies wrapped around an agent to stop it doing something unsafe, off-brand or non-compliant. On your side of the fence, the equivalent is deciding which actions an agent may take on your site at all.
Headless browser
A browser running without a visible interface, which is how most automated agents and rendering crawlers work. If you want to see what an agent sees, this is the honest way to look.
HTTP 402 Payment Required
A status code reserved since 1997 and left unused for most of that time, now the handshake for machine-to-machine payment, underneath both pay per crawl and the agent payment protocols. My favourite piece of trivia in either glossary.
Human-in-the-loop
Requiring a person to confirm before an agent does something irreversible, such as paying. It is standard practice in agentic commerce and in an agent readiness test a safe stop counts as a pass.
Human-present versus human-not-present
The payments industry’s framing for whether a person is actually there when a transaction happens. The rules for card-not-present transactions were written long before autonomous agents and the liability question follows from which category an agent purchase lands in.
Idempotency
Making a repeated request produce the same result as a single one, so nothing gets created twice. It matters because agents retry and a checkout that creates two orders on a single retry is a support ticket you will get to enjoy.
Indirect prompt injection
Malicious instructions hidden in content an agent reads, hijacking what it does next. Brave demonstrated screenshot-based invisible-text attacks against Perplexity Comet and a navigation-triggered attack against Fellou, concluding they all come down to a failure to separate trusted user input from untrusted web content. OpenAI has said publicly that it treats the problem as closer to evolving online scams than to a bug with a fix, which is the most honest thing anyone has said about it.
Instant Checkout
Scaled back. OpenAI’s in-ChatGPT purchase flow, built on the Agentic Commerce Protocol. It launched with US Etsy sellers in September 2025 and was scaled back in March 2026, with roughly a dozen of Shopify’s merchants live, single items only, no multi-item carts and incomplete tax handling. OpenAI’s own line was that the initial version lacked the flexibility it wanted. Keep this one in your back pocket for the next meeting where someone presents agentic commerce as inevitable and imminent.
llms.txt in an agentic context
Low adoption. The file has never been shown to lift AI search visibility and coding agents do fetch it. So its honest job is agent and developer onboarding. If you keep one, link it, keep it small and treat it as code, because a stale file that agents are built to trust is a security surface.
M to P
Machine-readable page variant
A stripped-down version of a page served to agents. Several vendors now sell it as a product. Adobe’s AI Content Visibility Checker scored US retail pages and found product pages the worst at 66 out of 100, with the weakest sites down at 54.2% against 82.5% for the best. So the problem it solves is real. See cloaking risk before you buy one.
Mandate
In AP2, the signed, verifiable artefact recording exactly what a user authorised an agent to buy and under what constraints. It is the mechanism that turns “my agent bought it” into something a merchant can check.
MCP (Model Context Protocol)
Production. An open protocol connecting AI applications to tools and data, usually a server you run exposing named tools an authorised agent can call. Anthropic handed it to the Agentic AI Foundation in December 2025 and it is now the closest thing this space has to a settled standard, with an actively versioned specification. Consider it when an authorised agent needs reliable access to a product, dataset or action that lives somewhere other than a public page.
Merchant of record
Who is legally selling and therefore who handles tax, refunds and disputes. Agentic checkout muddies it, because the customer transacted inside an assistant and the fulfilment happened somewhere else.
MPP (Machine Payments Protocol)
Early. Tempo and Stripe’s protocol for agents paying for API requests, tool calls and content over HTTP 402, with sessions for pay-as-you-go billing. Tempo is the associated chain, so this carries the same crypto dependency as x402. Newer and less established than ACP or AP2.
Multi-agent orchestration
Coordinating several specialised agents on parts of one job, so one researches while another writes. Mostly a builder’s concern and it turns up in vendor decks aimed at marketers often enough to be worth recognising.
NLWeb
Production. A project started at Microsoft for turning a website into a conversational, queryable interface agents can use, built on your existing Schema.org data and feeds. It reuses what your SEO team already maintains, which makes it cheaper to try than most things here.
OAI-SearchBot and ChatGPT-User
OpenAI’s two documented non-training tokens: OAI-SearchBot builds the index ChatGPT searches and ChatGPT-User is the user-triggered fetch, which is also what ChatGPT agent traffic arrives as. There is no separate OpenAI agent token. Operator had one and was shut down in August 2025, so ChatGPT-User is doing double duty and you cannot cleanly split agent activity from live browsing in OpenAI’s traffic.
OKF (Open Knowledge Format)
Proposal. A Google Cloud specification for representing reusable knowledge for agents, in practice a folder of cross-referenced markdown files with YAML frontmatter served from your site. Still a v0.1 draft. Monitor it.
Perplexity Comet
Production. Perplexity’s agentic browser, free across Windows, macOS, Android and iOS and the highest-share agentic browser in the measurements that exist at 47.6% of observed agentic traffic in HUMAN’s June 2026 census. It is also the subject of the most important legal case in this glossary, Amazon versus Perplexity.
Product feed for agents
Structured product data supplied directly to agentic surfaces, with current price, availability, variants and policies. Same discipline as Merchant Center feed hygiene, harder failure mode, because an agent that reads a stale price may complete a purchase on it.
Project Mariner
Discontinued. Google’s browser agent experiment, discontinued as a standalone Labs product on 4 May 2026, with its technology absorbed into Gemini Agent, AI Mode and Chrome. Worth knowing because it is still cited in articles as though it were live and because it is a useful example of how quickly this layer churns.
Prompt injection
Instructions hidden in content that hijack a model or agent reading it. The direct version comes through user input. See indirect prompt injection for the variant that arrives through your pages, which is the one publishers should care about.
Q to S
Rate limiting agents
Capping requests per interval. It protects your infrastructure and it can silently break an agent mid-task, which the customer experiences as your site being broken.
RFC 9421 (HTTP Message Signatures)
Production. The IETF standard for cryptographically signing HTTP requests, published in February 2024. It is the machinery underneath Web Bot Auth and the reason agent identity is becoming something you can verify.
robots.txt and agents
Your robots file governs crawlers. User-triggered fetchers such as Google-Agent and ChatGPT-User ignore it, on the logic that a person asked for the page. So a robots rule governs crawling and leaves agent behaviour untouched and assuming otherwise is common and expensive.
Sandboxing
Running an agent in an isolated environment so it cannot touch anything it was not given. Microsoft’s Copilot Actions runs in exactly this kind of agent workspace and it is the standard answer to the question of what happens when an agent goes wrong.
Schema.org for agents
The structured product, offer, price and availability markup your SEO team already maintains and the substrate almost everything else here sits on. NLWeb is built on it, UCP feeds restate it and an agent reads it long before it touches a protocol. If you do one thing on this page, make this accurate.
Selector stability
Whether the elements on your page keep the same identifiers between deploys. Agents that learn a journey break when class names are regenerated on every build, which is one of the few genuinely new technical asks agents make of a front end.
Semantic HTML
Correct headings, lists, tables, labels and landmarks, so a parser can work out structure without rendering anything. The cheapest agent readiness work there is and it was already on your list.
Session state
Whether a journey survives being paused, resumed and retried. Agents do all three and a basket that empties after 10 minutes fails a task a human would have finished.
Shared Payment Token
Production. A single-use, scoped payment credential passed to an agent in place of a real card number, so it can complete one purchase and nothing else. It is Stripe’s implementation of the Agentic Commerce Protocol’s delegated payment approach. It is how the commerce protocols intend to make agent purchases safe.
Shopify agentic commerce
Production. Shopify’s support for agentic surfaces. It matters disproportionately because so many merchants sit on the platform, so one decision there moves a large share of the market at once. OpenAI launched Instant Checkout with US Etsy sellers, with Shopify merchants following.
Signed agent
A bot or agent carrying verifiable cryptographic identity under Web Bot Auth, so an operator can allow, block or charge it by name. Anyone can copy a user agent string and nobody can copy a signature. Cloudflare and AWS have both shipped support.
Skills
Production. Packaged, reusable instruction sets that extend what an agent can do for a specific task, popularised by Anthropic’s SKILL.md format and now being proposed as an MCP extension. Relevant to marketers mainly as the mechanism by which an agent could be taught to work with your product properly.
Stripe agentic commerce tooling
Production. Stripe’s products for agent-initiated payment, including shared payment tokens and its co-authorship of the Agentic Commerce Protocol. Payments infrastructure is where much of this layer is being settled, a long way from where most marketing teams are looking.
Structured actions
Explicitly declaring what can be done on a page, so an agent is told which element adds to the basket and does not have to infer it from your markup. WebMCP is the browser-side version.
Subagent
A specialised agent spawned by another to handle one part of a task, with its own context. A builder’s term that you will meet in tooling discussions.
Terms of service
The control surface for agent access and after the Ninth Circuit narrowed computer-access law in Amazon versus Perplexity, the one that now decides who may automate against your site. If you have a view on agents, it belongs here as well as in robots.txt.
T to Z
Task completion rate
The share of attempted agent tasks finished correctly. It is the metric the benchmarks report and the one your site quietly affects, so if agentic traffic ever matters commercially this is the number you get judged on.
Task success rate
See task completion rate. Both names are in use, neither has a standard definition and vendors quoting one rarely say what counted as success.
Tokenised card credential
A payment credential that stands in for a real card number, limited in scope and revocable. The foundation of every safe-agent-purchase design currently proposed.
Tool calling
A model invoking an external function or API with structured arguments, which is the foundation of everything agentic and what MCP standardises.
Trust tier
A classification some networks and platforms are proposing, so a verified agent from a known operator gets treated differently from an unidentified one. It is where agent identity turns into commercial access.
Trusted agent
An agent an operator has chosen to recognise and allow, usually on the strength of a cryptographic signature. It is the practical output of agent identity work and what trust tiers are built on.
UCP (Universal Commerce Protocol)
Production. An open standard for agentic commerce, announced by Google in January 2026 with 20+ launch partners including Shopify, Etsy, Wayfair, Target and Walmart, compatible with AP2 and powering purchases in AI Mode. Google describes the protocol itself as neutral and vendor agnostic, while approving merchants before they go live on its own AI surfaces. This was the announcement that shifted my own reading of the space from isolated experiments to shared infrastructure. Standards being written is still a long way from standards being adopted.
Verifiable credentials
Production. A W3C standard for tamper-evident digital claims, now being applied to proving an agent holds the authority it claims. The generic answer to delegated authority, competing with the payment-specific mandate approach.
Visa Trusted Agent Protocol
Production. Visa’s specification for identifying agents in payment flows. It matters because a card network writing the rules carries a different kind of weight from a technology vendor proposing them.
Web Bot Auth
Draft standard. An emerging IETF standard where bots and agents cryptographically sign their requests with HTTP Message Signatures, so a site can verify who is calling instead of trusting a user agent string. Cloudflare and AWS have both shipped support. It is the most consequential item in either glossary that nobody in marketing is discussing yet.
WebMCP
Origin trial. JavaScript on your own pages declaring what a browser agent can do there. Chrome opened an experimental origin trial in June 2026 with Expedia, Booking.com, Shopify, TurboTax, Redfin, Etsy, Instacart and Target testing. Experiment with it once a browser task has failed through ordinary clicking and form interpretation and no earlier.
Well-known URI
The /.well-known/ path convention where machine-readable files live, so an agent can find them without being told. ARD publishes ai-catalog.json there and A2A publishes its agent card at /.well-known/agent-card.json. Knowing the convention is most of knowing where to put things.
x402
Production. A protocol from Coinbase reviving HTTP 402 for machine-to-machine micropayments over ordinary HTTP requests, settling in stablecoins on EVM chains. Worth knowing before a finance team asks. The x402 Foundation launched under the Linux Foundation in July 2026 with 40 member organisations, which moved it from interesting idea to governed standard in about a year.
The evidence worth knowing
Five findings that should shape how much of your quarter this gets.
Agents are good at looking and bad at doing
Web Bench ran roughly 2,450 tasks across 452 real websites. Five of the seven agents tested cleared 70% on read-only tasks. On the categories that change something, meaning create, update, delete and file manipulation, the best fully automated agent managed 46.6%. Every commercially interesting task is in the second group.
The most competitive agent completed 30% of realistic work tasks
Carnegie Mellon’s TheAgentCompany benchmark tested agents on consequential simulated workplace jobs and the most competitive one finished 30% of them. The “agents fail 76% of tasks” headline you will see quoted is the same metric on the same benchmark, run against Claude 3.5 Sonnet in December 2024, so it is an older model rather than a different framing.
Roughly a third of the average retail product page is unreadable to a model
Adobe’s AI Content Visibility Checker scored US retail pages: Homepages 75 out of 100, category pages 74, product pages 66. Across sites the spread runs from 82.5% at the best to 54.2% at the worst, so this is a site-quality problem more than a page-type one. Product pages, the ones that make money, still score worst. That is the single best argument for doing the boring work before the protocol work.
The traffic is small, growing fast and no longer low quality
Adobe measured AI traffic to US retail sites up 393% year on year in Q1 2026, converting 42% better than non-AI sources in March 2026 having converted 38% worse a year earlier. That is AI referral traffic, so humans arriving from an assistant and it belongs in a different column from agent activity.
An agent acting under human direction may legally be the human’s act
Amazon sued Perplexity over Comet shopping on its site and won a preliminary injunction. In August 2026 the Ninth Circuit overturned it, reasoning that it was the Comet user accessing Amazon’s computers rather than Perplexity, because the browser requires user direction. The underlying case continues. It is the first appellate signal on whether sites can block agents and anyone writing a bot policy this year should read it.
What I would actually do to start improving AI search accessibility now
Seven moves. The first two need nobody’s permission and the third is the one almost nobody runs.
- Run one agent test and write down where it fails: Pick a task that already carries commercial value, ask an agent to complete it on your live site and record the exact point it gives up. Claude for Chrome or Comet will do. Decide before you start what counts as success and keep a known-answer sheet of the correct product, price, availability and next action so you have something objective to compare against.
- Fix what the test found before you touch a protocol: It will be inaccessible content, inconsistent data, unclear form labels or a journey that breaks on retry. Rerun the identical task afterwards. In my experience this is where the gain is and no standard rescues a journey that does not work for a human either.
- Then test the protocols themselves, before and after: This is the step that gets skipped and it is cheap, because the knowledge-layer standards are just files. OKF is a folder of markdown files with YAML frontmatter. ARD is a single JSON manifest at
/.well-known/ai-catalog.json. EntityMap is two files at your domain root and it ships a validator so you can at least confirm the file is correct.llms.txtis one file. None of them needs a server, so you can publish one in an afternoon, send an agent at the identical task and compare against the baseline from step 1. Set your expectations honestly though. ARD is a v0.9 proposal and OKF a v0.1 draft and I know of no production agent fetching either from a public website, so expect no change. EntityMap is the exception worth watching, because its own vendor has published a 19-week before-and-after showing citations rising and moving toward product pages, on a single site with no control group. Publish these because they cost an afternoon and you want to be in place when consumption starts and measure anyway, because that is the only way anyone finds out. The execution-layer standards are a real commitment by contrast: A2A needs a live endpoint behind its JSON agent card, MCP needs a server you run, WebMCP needs JavaScript on your pages and is gated behind a Chrome origin trial and UCP needs a product feed Google approves first. I map the whole landscape and where each standard sits between knowledge and execution, in the AI Search Optimisation 101 guide. - Separate agents from crawlers in your logs: Google-Agent and ChatGPT-User are doing a different job from GPTBot and reporting them as one number hides the only interesting signal in the data. Note that OpenAI gives you no separable agent token, so Google-Agent is currently the cleaner signal of the two.
- Check your feed, not your copy: An agent reading a stale price may complete a purchase on it. Prices, availability, variants and policies first.
- Decide your agent policy deliberately: Blocking a shopping agent blocks a buyer, which is why this decision belongs with commercial owners and not with whoever happens to manage the WAF.
- Watch for the defaults: Agentic capability arriving switched on in a major browser or operating system is the signal that changes the priority and it has not happened yet.
Run steps 1 to 3 as one loop and you have something almost nobody in this conversation has: A measured before and after on your own site, rather than a vendor’s benchmark.
Agentic search FAQs
What is agentic search optimisation?
Making a website something an AI agent can act on, all the way through to completing a task. It covers the technical basics an agent depends on, structured product and pricing data, forms and journeys that survive automation and the protocols that let an agent discover capabilities or make a payment. It sits alongside AI search optimisation, which is about being found and cited in the first place.
Do I need to implement MCP, A2A or WebMCP?
Run one real agent task on your site first, because most failures turn out to be ordinary website failures that no protocol will rescue. After that, separate the cheap standards from the expensive ones. OKF, ARD, EntityMap and llms.txt are static markdown or JSON files you can publish in an afternoon, so there is little reason not to test them. Expect no measurable lift from most of them and measure anyway. MCP, A2A, WebMCP and UCP need running code, a live endpoint or an approved feed, so those wait until a structural gap survives the fixes.
Should I block AI agents from my website?
Treat it as a commercial decision first and an infrastructure one second, because an agent arrives when a specific person asked it to do something, so blocking it can mean blocking a buyer. Worth knowing that in August 2026 the Ninth Circuit overturned Amazon’s injunction against Perplexity’s Comet on the reasoning that the user, not the browser maker, was accessing Amazon.
How do I measure agent traffic?
With difficulty. Agent sessions look human in analytics, trigger pixels and remarketing and blur attribution and there is no shared standard for identifying them. Server logs are the most reliable source available today, where a named agent user agent such as Google-Agent can be separated from crawlers. OpenAI publishes no separate agent token, so its agent traffic arrives as ChatGPT-User alongside ordinary live browsing.
Is agentic commerce actually happening?
It is real, small and volatile. Adobe measured AI traffic to US retail up 393% year on year in Q1 2026 with conversion quality improving sharply and OpenAI scaled back Instant Checkout in March 2026 after launching with about a dozen merchants, single items only and no multi-item carts. Both things are true and anyone showing you only one of them is selling something.
Work with me on agentic search
Knowing the vocabulary is the easy part. If you want to know which of these terms is currently costing you money - crawlers that cannot reach the pages that matter, an AI answer describing your brand wrongly, or a visibility number nobody can defend - tell me what you are seeing and I will tell you whether it is worth fixing.