Section 1
The five challenges at a glance
The technical gap between a typical service-business website and an agent-readable one is wider than most founders assume, because the site was built to persuade humans and is now being read by parsers. The table below isolates five failure modes. The first is extraction failure: visually rich, JavaScript-heavy pages that render beautifully for people and yield noise for machines. The second is the absence of any curated machine entry point, the problem llms.txt was proposed to solve (Answer.AI, 2024). The third is silent schema exclusion: without typed Service, Offer and Review data, agents have nothing to compare, so the firm is dropped from shortlists without notification. The fourth is transactional dead ends: an agent that can research you but cannot book, quote or buy hands the completed sale to a competitor that exposed an endpoint. The fifth is crawler-policy chaos: robots.txt files that either block the AI crawlers a firm depends on for visibility or admit scrapers it intended to refuse. Every one of these failures is invisible in normal analytics, which is what makes the category dangerous: traffic dashboards show humans, not the agents that never sent them. The remedy in each case is a published, open standard rather than proprietary tooling, the rare situation in technology where the right answer is also the cheap one.
Section 2
Schema.org: the oldest and most proven layer
Schema.org is the foundation because it is the standard with the longest track record and the broadest consumption. Founded in 2011 by Google, Microsoft, Yahoo and Yandex as a shared vocabulary for structured data on the web, it lets a site declare typed facts, this is an Organization, this is a Service, this Offer has this price, most commonly in JSON-LD blocks embedded in the page (Schema.org, 2024). For service businesses the relevant types are Organization (identity, sameAs links to corroborating profiles), LocalBusiness where applicable, Service and Offer (what is sold and on what terms), FAQPage (direct answers agents can lift), and Review with AggregateRating (third-party proof). Two honest qualifications. First, schema.org was built for search engines, and its guaranteed consumers remain search features; AI assistants and shopping agents increasingly draw on the same structured feeds, and agentic-commerce implementations explicitly request rich product and offer data, but no LLM vendor contractually promises to read your markup. Second, markup is a claim, not a verification, which is why the sameAs corroboration graph matters: agents cross-check declared facts against directories, review platforms and registries, and consistency is itself a trust signal (Digital Commerce 360, 2025). Implementation discipline beats volume. Five accurate, validated types outperform fifty sloppy ones, and contradictions between markup and visible page content are the kind of discrepancy a cautious agent treats as disqualifying. Validate everything against schema.org and search-console tooling quarterly.
Section 3
llms.txt: a useful proposal, honestly labeled
In September 2024, Jeremy Howard, co-founder of Answer.AI, proposed llms.txt: a markdown file at a site's root that gives language models a curated, token-efficient guide to the site's most important content (Answer.AI, 2024). The rationale is practical, context windows are finite, and HTML pages stuffed with navigation, scripts and ads convert poorly into model-usable text, so the file offers a clean map: a title, a summary, then annotated links organized by section, with an optional tier that can be skipped when context is tight (Answer.AI, 2024). The honest status report: llms.txt is a proposed standard, not a ratified one. Documentation-heavy companies, developer-tool firms prominent among them, have adopted it, and tooling ecosystems generate it automatically, but the major model providers have not formally committed to consuming it, and skeptics inside the search community note the absence of confirmed crawler support. A service business should hold both facts simultaneously: unproven consumption, near-zero cost. The file takes an afternoon, doubles as a forcing function, writing it makes you decide which ten pages actually define the firm, and positions you for whichever consumption pattern emerges. What it does not do is substitute for schema.org markup or quality content; treat it as the index, not the library. Pair it with a deliberate robots.txt policy for AI crawlers: decide explicitly which agents you admit, because visibility in assistants requires being crawlable by them (Gartner, 2024).
Section 4
APIs and transaction protocols: where readability becomes revenue
Reading is the prerequisite; transacting is the prize, and the standards here moved from paper to production in 2025. The Agentic Commerce Protocol (ACP), co-developed by OpenAI and Stripe and released as an open standard under the Apache 2.0 license, defines how an AI agent completes a purchase on a buyer's behalf: checkout sessions, cart management, delegated payment tokens that pass between buyer, agent and business, and OAuth-based delegated authentication (Stripe, 2025). It is live, not theoretical, ACP powers Instant Checkout in ChatGPT, beginning with US Etsy sellers and extending to more than a million Shopify merchants (OpenAI, 2025), and Stripe states that existing Stripe businesses can enable agentic payments with minimal integration changes (Stripe, 2025). Alongside it sits Anthropic's Model Context Protocol (MCP), an open standard introduced in November 2024 for connecting AI systems to tools and data sources (Anthropic, 2024); where ACP standardizes the purchase, MCP standardizes the plumbing agents use to query systems, calendars, catalogs, CRMs. For service businesses the near-term translation is concrete rather than speculative: expose a bookable surface. An online scheduling system with a clean API, a productized offer purchasable through a standard payment link, an intake form an agent can complete, each is a primitive the protocols are converging on. Retail is the proving ground today; Gartner's $15 trillion B2B intermediation forecast (Gartner, 2025) is the argument that B2B services are next.
Section 5
Innovative solutions
Beyond the baseline standards, early adopters are layering practices that compound. The first is GEO-informed content engineering: the KDD 2024 study showed that citing sources, adding statistics and including quotations lifted generative-engine visibility by up to 40%, with domain-dependent variation (Aggarwal et al., 2024), so leading firms now write key pages as evidence artifacts, every claim sourced, every number dated. The second is the machine-readable proof page: a single canonical URL aggregating credentials, named results with figures, and review links, marked up with schema and referenced in llms.txt, built explicitly to be the page an agent cites. The third is API-first intake: replacing 'contact us' with structured intake, budget range, timeline, scope selectors, that both humans and agents can complete, shrinking qualification cycles for everyone. The fourth is corroboration management: systematically aligning name, address, pricing and claims across every third-party surface an agent might cross-check, treating inconsistency as the trust leak Gartner's verifiable-data framing implies it is (Digital Commerce 360, 2025). The fifth is monitoring: monthly citation audits across ChatGPT, Claude, Perplexity and Gemini, logging presence, accuracy and cited sources, because agent-channel visibility appears in no standard analytics dashboard. None of this requires engineering headcount; all of it requires editorial and operational discipline, which is precisely why it differentiates: budgets can buy ads, but verifiable consistency has to be operated.
Section 6
Solution framework
Sequence the build as four layers, each a precondition for the next. Layer one, identity: accurate Organization schema with sameAs links, a deliberate robots.txt policy for AI crawlers, and llms.txt mapping your defining pages. This layer answers 'who are you' consistently everywhere a machine might ask. Layer two, offering: productized services expressed as Service and Offer types with scope, deliverables and price bands; FAQPage markup carrying direct answers to the questions buyers delegate to assistants. This layer answers 'what do you sell and on what terms.' Layer three, proof: Review and AggregateRating markup, numbered case studies, third-party profiles kept consistent with on-site claims. This layer answers 'why should an agent rank you,' and it is where the GEO evidence applies most directly, sourced statistics and quotations are the currency (Aggarwal et al., 2024). Layer four, transaction: a bookable or buyable endpoint, ACP-compatible payment infrastructure where applicable (Stripe, 2025), MCP-aware tooling as your systems mature (Anthropic, 2024). This layer converts readability into revenue. The framework deliberately inverts the typical founder instinct, which is to start with the exciting transactional layer; an agent that can buy from a firm it cannot verify will not buy from it. Audit quarterly, treat every layer as production infrastructure with an owner, and version your llms.txt and schema the way you version code.
Section 7
Evidence-based action plan
Days 1-30: audit and policy. Run your key pages through schema validation tooling; fetch your own site the way a crawler does and inspect what survives; review robots.txt against the current roster of AI user agents and set an explicit allow policy; baseline your citation presence across the four major assistants. Days 31-60: implement the read layer. Deploy Organization, Service, Offer, FAQPage and Review markup in JSON-LD; publish llms.txt with your ten defining pages annotated; rewrite your two highest-intent pages with sourced statistics and quotations per the KDD findings (Aggarwal et al., 2024); align every third-party profile with on-site facts. Days 61-90: implement the transact layer. Stand up structured intake; make at least one productized offer bookable or purchasable end to end; if you process payments with Stripe, review ACP applicability now rather than later (Stripe, 2025). Then operationalize: monthly assistant-citation audits, quarterly schema validation, and an owner for machine readability named in your operating cadence. Honest framing for the board: schema.org is proven infrastructure; GEO tactics carry peer-reviewed evidence; ACP is live but retail-first; llms.txt is a low-cost bet on an unratified proposal; and the $15 trillion that justifies the urgency is a Gartner prediction (Gartner, 2025). The portfolio logic holds because the downside of every action is a better-documented business, and the upside is being legible to the buyers, human and machine, that the next decade sends. For adjacent evidence in this pillar, see [The Agent-Washing Problem: How to Evaluate AI Agent Vendors Without Getting Burned](/blog/growth-agent-washing-vendor-evaluation) and [Human Relationships in an Agentic Market: The Evidence for Pricing the Human Layer](/blog/growth-human-relationships-agentic-market).