Computational Marketing

Vertical AI-marketing playbooks are recipes. The asset underneath is a site an agent can read, check and book through

Copying a competitor's playbook fails if agents can't use your site.

· 8 min read
Cover illustration for “Vertical AI-marketing playbooks are recipes. The asset underneath is a site an agent can read, check and book through”
Features · November 18, 2026 · 8 min read · 1,776 words

Vertical AI-marketing playbooks are recipes. Whether a hospitality GEO playbook or another industry guide works depends on the site beneath it. The lasting asset is a site an agent can read, check and book through. Before they choose a tactic, technical marketers need to ask whether the site’s architecture lets it work.

Why playbooks circulate but rarely deliver consistent results

A competitor publishes a “hospitality GEO playbook,” and the pressure to replicate it is easy to understand. It gives marketers a clear way to respond as search changes. But what, exactly, is being copied? A playbook records what worked for another site, in another setting, with an existing architecture that the recipe may barely describe. Following its instructions does not reproduce those conditions.

That distinction becomes sharper when every competitor adopts the same FAQ structure, schema types and entity-seeding checklist. What remains is the site each brand brought to it, including its structure, factual consistency and machine-readability. Generative engine optimization depends on whether content can be retrieved, is fit for re-ranking and is worth citing in an answer. Those are properties of the published site.

What AI answer engines actually evaluate when they decide what to cite

An answer engine works with discrete signals it can retrieve. Retrieval-augmented generation systems, used by Perplexity, Google AI Overviews and others, retrieve pages at query time. If a page cannot be crawled or parsed, its topical ambition cannot make it available for citation. Access is the first constraint.

Once content is accessible, the question becomes whether the engine can form a coherent representation from it. Conflicting facts, inconsistent names and mismatched data across pages weaken that representation, even when one page has been carefully optimized. Can the engine understand the claim, verify it and quote it? These are separate tests. Each depends on how information is organized and connected across the site.

How the answer-engine layer is giving way to an action-engine layer

Google began rolling out hotel booking inside AI Mode in the U.S. on August 27, 2026, with launch partners including Booking.com, Expedia, Hilton, IHG and Marriott. The hotel example changes the endpoint of search. An answer engine can stop at recommending a property, while an action engine works toward a booking confirmation.

Claude for Chrome also shows how agents can operate on live sites. It can see page content, click buttons and fill forms, with site-level permissions and confirmation prompts before risky actions. Reading is only part of the task. An unreadable form, a JavaScript-gated price or a session requiring human navigation can prevent completion after a brand has already been found. In that setting, the competitive question is not only which property gets recommended, but which property the agent can successfully book.

The two distinct ways a site can fail in an agentic environment

Citation failure happens when an engine cannot retrieve, parse or represent a brand accurately enough to include it in an answer. Content structure, crawlability, entity consistency and external signals address this problem. Transaction failure happens after an agent has already found the brand. An agent may find and cite a hotel but still be unable to complete its seven-step booking form, reach a price behind a login or interact with its calendar widget. Improving the cited page does not necessarily repair the booking path.

Entity inconsistency sits between these two failures as a distinct verification problem. An agent cross-referencing conflicting room names, check-in times or amenity lists cannot assemble a reliable account of the property. Contradictory structured data can be worse than missing data because it undermines confidence in the source. Hilton’s FY2025 10-K names large language models entering internet travel bookings as a risk that could divert bookings from direct channels.

What entity consistency actually requires across a site

Entity consistency means that the same thing has the same name and facts wherever an engine encounters it. That sounds editorial, but its enforcement is architectural. A product called “Advanced Plan” on a pricing page should not become “Pro Tier” in a blog post and “Enterprise Starter” in a knowledge base article. Each variation fragments the signal the engine is trying to assemble. In hospitality, the equivalent is a room or amenity described differently on the brand site, Google Business Profile and OTA listings.

Those facts need governance through the CMS and data layer. Controlled vocabularies and canonical names provide the common reference. Organization, Product and FAQ schema applied at the template level can then propagate consistent entity signals across generated pages. Markup is not the whole discipline. The underlying names and claims must agree so that structured data and visible content describe the same business.

Why crawlability and bot permissions are strategic decisions, not IT housekeeping

Bot access can decide whether an otherwise usable booking path is available at all. The October 2026 hotel-agent test found that checkout access was often decided by bot settings. A CDN rule or an old robots.txt entry could block the booking path even when the content elsewhere was strong. That makes access a commercial decision.

Different AI crawlers also serve different purposes. GPTBot controls ChatGPT’s training inclusion, while OAI-SearchBot controls live retrieval for ChatGPT search. ClaudeBot serves Anthropic, and PerplexityBot serves Perplexity. A rule written to manage Google’s crawler does not automatically express the right policy for these identities.

Prices, availability and forms that depend on JavaScript are unreadable to crawlers and agents that do not execute it, while a page that times out during traversal is unavailable. Growth and marketing leaders should review bot permissions with engineering to check which crawler identities can reach which types of paths. An old configuration is not a strategy. The review needs to connect infrastructure choices to discovery and booking access.

The off-site entity graph: earning citations on surfaces AI engines already trust

A brand’s own domain is only part of its retrievable identity. AI engines also learn from and retrieve the broader web. Chen et al.’s 2025 paper found that AI engines strongly favor earned media from authoritative third-party sources over a brand’s own content. A well-structured site can therefore remain disadvantaged if the brand is absent from the sources engines retrieve for its category. External presence is part of the evidence available to the engine.

Rise rebuilt content around citable, expert-led explanations and secured external mentions across trade press and analyst notes, producing substantial organic growth and expanded AI-search visibility. No special AI-specific page type was involved.

For technical marketers, the corresponding method starts with the prompts buyers actually use and the sources cited in response. Substantive content and earned mentions can then be placed within that source graph, including industry publications, analyst notes, review platforms and trade press. Consistency still governs the work. External placements need to reinforce the same names and claims as the brand’s domain, rather than create a separate version of the business.

Which tactical shortcuts the evidence does not support

Dedicated pages labeled “for AI assistants” have no evidence of receiving different treatment from retrieval systems. A crawlable, well-structured site already supplies what those pages attempt to signal. The label does not establish usefulness or trust.

Google Search’s May 15, 2026 guidance states that llms.txt is not needed for AI Overviews, AI Mode or any other generative AI Search feature. It groups that tactic with content chunking, AI-specific rewriting and special schema as approaches that do not help Google AI visibility. This guidance distinguishes ordinary content structure and consistent structured data as foundations from special treatments promoted as shortcuts to selection. Tactics that promise durable gains in a day without addressing retrieval-time relevance, entity coherence, content structure or domain authority deserve scrutiny.

How to evaluate any playbook tactic against the site's actual architecture

Technical marketers should begin with the buyer’s prompt, then audit the architecture, and only then choose the tactic. Does the site’s architecture permit this tactic to function? If the answer is no, the structural repair comes first. Otherwise, the team is paying for work before the conditions it needs are in place.

For a citation tactic, those conditions include unblocked content paths, core content readable without JavaScript execution, an extractable heading hierarchy and consistent entity terminology. A transaction tactic needs a different foundation. Booking and form paths must let the relevant agent user-agents in, while pricing, availability and service definitions must be readable without a human navigating the site. Confirmation flows must also work without requiring session-state that the agent cannot maintain. Success in the citation layer is not evidence that these conditions hold.

Entity-seeding work has its own prerequisite, namely canonical names and factual claims defined at the data layer and propagated through templates, profiles and earned placements. Without that reference, greater distribution can spread contradictions. Measurement should then track citation frequency across a defined prompt set by engine, including ChatGPT, Claude, Gemini and Perplexity, as well as citation accuracy. AI referral traffic offers a directional signal. “Share of answer” across the prompt set, rather than organic rank, is the primary KPI for judging whether the brand is appearing in the answers buyers receive.

What a site built for the agentic layer looks like in practice

A site built for agents does not need to look like a different species of marketing site. Clean headings create extractable passages, server-rendered or static content exposes essential facts without JavaScript, and bot permissions reflect deliberate choices about each crawler and path. Canonical names and facts flow from the data layer into pages and external profiles. That identity can be checked across sources without resolving avoidable contradictions.

The same clarity extends into the transaction. Booking, quote and inquiry paths are accessible to agents, with readable pricing, availability and service definitions. Agents can complete forms without session barriers they cannot handle, and confirmation endpoints show in machine-readable form that the task is done. Content supports that journey by answering the query in its first passage before expanding. Conversational FAQs and original, citable claims give the engine both a usable answer and information it cannot find everywhere else.

Maintaining that system is an ongoing production task. Zapier’s multi-agent content pipeline illustrates how a designed workflow can combine scale with review. Researchers feed writers, then four specialized editor agents address customer language, style, AI-generated phrasing and relevant links before human review. A standing prompt tracker then checks the published work for citation frequency, sentiment drift and accuracy, because producing more content does not in itself show that the content is being seen.

That is the asset a playbook can build upon. Each new page, external mention and transaction path that works adds to a site that keeps the same identity as it grows. A competitor can copy the recipe without acquiring that foundation.

More in Features