The Agency Stack for AI Search in 2026
The measurement, crawler, analytics, reporting, and workflow layers we use to run accountable AI search programs for clients.
An agency AI search stack should answer five questions. What did the model say, which source did it use, could a crawler read the client's page, did a person visit, and what changed after the team acted?
A separate dashboard for each question creates more reconciliation than insight. We use stable identifiers across systems while keeping the prompt, response, cited URL, crawler request, visit, and conversion as distinct records.
Our client programs use Promptwatch as the AI visibility layer beside the client's existing CMS, search, analytics, edge, reporting, and work-management systems. We do not replace a working content operation to make the tooling look tidy.
The website remains the source of truth
The production site comes first. Its CMS owns claims, dates, canonical URLs, authorship, and page status. The sitemap lists canonical pages, while structured data describes visible facts and redirects preserve route changes.
Much of this is technical SEO. AI measurement cannot compensate for an outdated sitemap, empty crawler HTML, or contradictory product facts. Someone must be able to update the page and approve sensitive claims.
A page change log records publications, substantial edits, migrations, and access-policy changes. It gives later charts context and discourages invented causal stories.
Search Console supplies the classic search view
Google Search Console remains the source for Google's search performance and indexing observations. Promptwatch's Google Search Console integration imports clicks, impressions, click-through rate, average position, page inventory, and URL Inspection status. The integration page says access is read-only and can pull up to a year of performance history.
We use those data beside AI visibility, not as a substitute for it. Search Console does not report ChatGPT or Perplexity responses. A Google query can seed prompt research, but a query impression and an AI mention are different events.
The page inventory lets us compare what Google knows with the sitemap and AI crawler requests. A URL missing from one list raises a question, not proof of an indexing failure everywhere.
Promptwatch holds prompts, answers, and sources
For ongoing measurement, we need a fixed set of buyer questions grouped by market and intent. Promptwatch records responses across selected AI platforms, detects brand mentions, stores citations, and tracks trends over time. We use the response itself as the evidence behind every aggregate.
The platform also separates mentions from citations. A brand can be discussed without its domain being used as a source. Its page can also support an answer that never names the brand. Those states lead to different work, so the schema for our reporting preserves both.
Page-level tracking follows published and optimized URLs against monitored prompts. Content gap analysis helps compare those prompts with indexed site coverage. These features guide review; they do not decide what a client should claim or guarantee that a page will be cited.
The practical reason we run this layer in Promptwatch is continuity. The same project can connect prompts, citations, crawler logs, visitor analytics, content actions, and reports. Fewer joins mean fewer chances to map a URL or market incorrectly.
Edge logs explain crawler access
Ordinary web analytics often omit crawler requests. Edge data can reveal denials that occurred before a request reached the application.
For Cloudflare clients, Promptwatch's Cloudflare integration documents two connection paths: Logpush for Enterprise and a Worker path for other plans. It sends timestamps, URLs, status codes, and user agents into crawler analytics. The page also says the product separates training crawls, search indexing requests, and citation fetches.
We evaluate the client's actual infrastructure rather than moving them to Cloudflare for the sake of one connector. The relevant requirement is reliable request data at the edge, with enough context to distinguish a 200 from a 403 or 404. Access credentials should be scoped, stored through the client's normal controls, and removed when setup is complete.
Crawler counts are diagnostic. A successful request means the bot reached a URL. It does not mean the page entered an answer. We connect the log to page and citation records so the team can see where the chain stops.
Analytics records visits and conversions
Referral and conversion data belongs in the client's approved analytics system. Promptwatch can also collect AI-referred visitor events and conversions through its visitor analytics setup. We keep channel definitions documented so a visit is not counted differently in two reports without explanation.
Attribution needs restraint. A referral from an AI platform is observable when the request carries the relevant source information. A later direct visit may have been influenced by an earlier answer, but the direct session does not prove that. We report tracked conversions and leave unobserved influence unassigned.
This layer is also where consent, retention, and data-access policies apply. GEO does not create an exemption from the client's analytics governance.
Looker Studio carries the client report
Most clients do not need another login for a monthly review. Promptwatch's Looker Studio connector exposes visibility, monitors, citations, sentiment, prompts, and crawler data through an API-key connection. Its integration page says reports update without repeated spreadsheet exports.
Our executive view includes observed referrals and conversions where available, mention and citation movement on a fixed prompt set, important model or topic cuts, top owned source pages, and unresolved access faults. Supporting pages hold response text and detailed prompt diagnostics.
We use separate charts rather than a synthetic score that blends traffic, sentiment, crawls, and citations. The report should tell an owner what to do next, not merely produce a number that always moves.
Slack and work management turn findings into work
An insight becomes useful when it reaches the person who can act. Promptwatch's Slack integration supports project mapping, questions against live data, and scheduled briefings. We use channel delivery selectively. Routine noise trains people to ignore the signal.
The actual ticket still belongs in the client's work system. Crawler blocks go to infrastructure. Missing or weak passages go to content. Inaccurate brand framing may need communications, product, or legal review. Each ticket links back to the prompt, response, URL, and date that justified it.
Access is role-based and reviewed when team members or agency partners change. API keys used for reporting are separated from credentials used for operational actions.
The stack earns its place through decisions
We audit the stack quarterly by asking which fields drove action. Prompts that never inform a decision can be retired. Reports nobody opens should be simplified. Duplicate analytics pipelines should be reconciled or removed. The point is a dependable operating loop, not maximum software coverage.
The minimum viable setup is a representative prompt baseline, citation-level evidence, crawler visibility for priority pages, referral measurement, and a route from finding to ticket. Add automation only after those records agree. That gives an agency enough to diagnose absence, document work, and report results without turning a collection of integrations into a claim the data cannot support.