Level 1 is the proven part: get named when buyers ask ChatGPT and Google's AI (copy of AEO Engine, $175.5k a month verified). Then we make the business readable, then operable in a browser, then callable by agents, then trusted to act, each with safeguards.
Start where the money is proven: being named in AI answers. Then make the business something agents can read, use and book. You stop being only a website and become an API an agent can call safely.
Named when buyers ask ChatGPT, Gemini, Perplexity, Google's AI. The retainer: $1,500 / $2,900 a month.
proven: AEO Engine $175.5k MRR (Stripe-verified)→Phase 2Prices, availability and policies in plain HTML and data; signed agents let through; forms an agent can fill.
2–5 days of work; bundled into the retainer→Phase 3An MCP server / booking API with read tools, capped write tools and human approval above limits.
$3,500–4,500 build + $400–750 a monthWhat's proven and what isn't. - Level 1 (being named by AI) has verified money: AEO Engine at $175.5k a month. - Levels 2–3 (agent-readable pages, agent-operable forms) are cheap fixes that agents already need. ChatGPT agent, Claude in Chrome, Comet and Google-Agent all work through websites today. - Level 4 (the business as an API) has a real price market but no verified small seller yet. Published MCP builds run $8–50k, and a Ukrainian studio quotes $3.75–9k for a ChatGPT app. - Level 5 (agents paying without approval) runs through payment networks, not our code. Visa's agent payments went live in Europe on 2 Jul 2026; Mastercard is still in pilots.
The order we sell in. Lead with level 1. Bundle levels 2–3 into the retainer. Sell level 4 to software firms first and run a booking pilot in the villa network.
Warning sign from the market. ChatGPT Instant Checkout launched Sep 2025 and ended Mar 2026 with ~12–30 merchants. Shopify now switches agent channels on for free. So we sell data quality and safe tools, not "connections".
Each level builds on the one below. Status as of Oct 2026 (live / beta / announced), from the platforms' own docs.
When someone asks ChatGPT, Gemini or Perplexity for the best X, the business is named in the answer.
robots.txt, HTML, sitemaps / IndexNow
A robot can read the price, what's included, availability, delivery, returns and terms as plain text, without guessing or clicking.
schema.org, merchant listings, feeds, NLWeb, llms.txt (optional)
The user's agent opens the site, fills the quote or booking form and reaches checkout without getting stuck.
Web Bot Auth (RFC 9421 signatures), WebMCP (Chrome trial), accessibility tree
Instead of clicking, the agent calls get_quote, check_availability, create_booking and gets exact, binding answers back as data.
MCP, OpenAPI, Google's UCP, A2A, ChatGPT apps
The user's agent pays a deposit or buys, and the business accepts automatically: the agent is identified, consent is provable, the amount is capped.
AP2 mandates, ACP delegated payments, UCP payment handlers, Visa TAP, Mastercard Agent Pay
Adoption numbers: - The official MCP registry lists ~51,100 servers. We paged its public API in full on 27 Oct 2026; the count includes duplicates and test entries. - Anthropic reported 10,000+ active public MCP servers in Dec 2025. - AI-agent requests ran at ~25B in Q3 2026, +41.5% on the quarter (DataDome via Help Net Security). - Shopify reports AI traffic to its stores up 8× year on year in Q1 2026 (Shopify). - Cloudflare blocks AI crawlers by default on new domains since 1 Jul 2025, and folded signed agents into its verified bots on 1 Jul 2026.
Protocols in one line each: - MCP: how agents call tools. - UCP: Google + Shopify's commerce protocol (catalog, checkout, a draft lodging spec). - ACP: OpenAI + Stripe's checkout and delegated-payment spec. - AP2: Google's payment mandates. - A2A: agent-to-agent messaging. - Web Bot Auth: agents sign their requests so sites can verify them. - Visa TAP / Mastercard Agent Pay: card-network agent identity and payments.
The ladder levels are our framing; every status is from the platform's own page.
The same task on two versions of a site. Left: what most sites look like today. Right: what we build. Illustrative examples; the rules behind them are from the platforms' docs.
An agent can't pass a CAPTCHA or read a PDF price list. So it books your competitor instead.
Outcome: The lead goes to a competitor whose price the agent could read.
Outcome: About 5 steps from question to running trial.
Outcome: The agent hands back a link and the purchase stalls.
Outcome: A sale the agent could finish, or a one-tap checkout.
Outcome: The booking goes to a booking platform that takes the commission.
Outcome: A direct booking, no OTA commission, in about 6 tool calls.
Why sites fail agents today. Cloudflare (about 20% of the web) blocks AI crawlers by default on new domains since 1 Jul 2025. Many sites also allowlist bots by user-agent string, which anyone can fake. The fix is to verify signed agents (ChatGPT signs with `Signature-Agent: "https://chatgpt.com"`) and rate-limit instead of challenging.
Sources: - Cloudflare default block - Cloudflare signed agents - ChatGPT agent allowlisting - Google merchant listing markup - Shopify agents - Stripe booking for agents - UCP lodging draft
The example sites are illustrative. Prices and names are made up; the behaviours come from the docs above and our research notes.
The business becomes something an agent can call, like an API, with limits we set. Copied from offers with published market prices; sold as add-ons to the visibility retainer, because the proven money is still in visibility.
Not "a service to sell" but an exposed API an agent can use: read anything, act within limits, escalate to a human above them.
An MCP server plus a ChatGPT app on the client's public docs and API, listed in the official MCP registry and the ChatGPT directory. A buyer's assistant can compare plans, check integrations and start a trial without a sales call.
Tools the agent can callget_pricingcompare_planssearch_docsget_integrationscheck_statusbook_demostart_trialProven price market: published MCP build agencies (AppMatic $8k–25k) and a Ukrainian AI-first studio quoting ChatGPT apps at $3.75–9k. No verified revenue for a small seller yet.
Make the catalog read well to shopping agents: Shopify's agent channels (on for free since 24 Mar 2026), Merchant Center feed, complete product facts, shipping and returns answers, monthly agent shopping tests.
Tools the agent can callsearch_productscheck_stocksize_adviceshipping_etaorder_statusstart_returncart → Shopify CheckoutChatGPT's own in-chat checkout (Instant Checkout) ended in Mar 2026 with ~12–30 merchants live; Shopify now gives agent channels away free, so we sell the data quality, not the connection.
A quote / availability / booking endpoint on the client's calendar or property system, following Stripe's booking pattern and Google's draft lodging spec. An agent can hold a slot; a human confirms.
Tools the agent can callsearch_availabilityget_quoteget_bookingcreate_holdrequest_changerequest_cancelAgent payments for bookings need the client's own Stripe in a supported country (UK, IE, EU, US, CA); Indonesia and Thailand aren't listed, so villas start with human-confirmed holds.
read read-only toolwrite tool that changes something (needs limits)Who already sells pieces of this (with published prices): - Mintlify: MCP server included in its docs plans, from $450 a month. - Stainless: MCP servers generated from API specs, free tier. - Speakeasy: agent gateway, pay as you go. - Zapier MCP: 500,000+ MCP servers created (self-reported). - AppMatic: MCP builds from $8k. - dForce: ChatGPT apps at $3.75–9k. - Goodie: free 36-check agent-readiness audit as a lead magnet. - Agent-readiness scanners sell for $9–99. That's a commodity, so we give it away in the snapshot.
What didn't work: - Smithery shut down its MCP hosting layer. Pure hosting is hard to charge for. - ChatGPT Instant Checkout ended in Mar 2026.
Our build stack: - AI generates the MCP server from the client's OpenAPI spec, plus tests and the app manifest. - Hosted in the client's cloud or on a Cloudflare Worker. - The language model behind any client-facing tool is not Anthropic's API (your billing rule); most tools need no language model at all.
A $10k mix (estimate): 5 visibility retainers + 4 run fees ≈ $9.8k a month.
Agents can use the business without anyone approving each request, because every call passes the same checks, and anything above a limit goes to a human.
Read anything. Write within limits. Above the limit, a human. Everything logged.
Verify the signed agent (Web Bot Auth; Visa TAP for payments). Never trust the user-agent string.
identityOAuth scopes, least privilege. Read tools open; write tools need a scoped token.
scopesAmount, nights, seats, per-day caps; quotes and holds expire; Idempotency-Key on every write.
limits"Needs a human" + a link to a prefilled human checkout or an owner approval.
escalateScoped payment token with max amount and expiry; the business never sees card data.
paymentsEvery call logged without personal data; cancel/refund tools; kill switch.
audit| Safeguard | What it does | Needed from level | Standard / source |
|---|---|---|---|
| Read vs write tools | Separate read tools from state-changing tools; truthful readOnlyHint/destructiveHint/openWorldHint (required in ChatGPT review) | L4 | developers.openai.com |
| OAuth 2.1 least-privilege scopes | Minimal initial scope, step-up via WWW-Authenticate scope challenge; no wildcard scopes | L4 | modelcontextprotocol.io |
| Token audience validation | 'MCP servers MUST NOT accept any tokens that were not explicitly issued for the MCP server' (no token passthrough) | L4 | modelcontextprotocol.io |
| Per-client consent (confused deputy) | Proxy servers MUST keep per-client consent, exact redirect_uri match, single-use state | L4 | modelcontextprotocol.io |
| State handle binding | Cart/booking handles random and bound to the authenticated user; possession is not authentication | L4 | modelcontextprotocol.io |
| Validate, rate-limit, sanitise | Servers MUST validate inputs, enforce access controls, rate limit, sanitise outputs | L4 | modelcontextprotocol.io |
| Human-in-the-loop thresholds | Human able to deny tool calls; confirmation for irreversible operations; business sets auto-accept bands (amount, lead time, party size) | L4 | developers.openai.com |
| Spending/booking limits on credential | Stripe SPT usage_limits: currency, max_amount, expires_at (<=180 days); Visa consumer-defined parameters | L5 | docs.stripe.com |
| Idempotency keys | Accept Idempotency-Key; same key+payload returns original; different payload 409; keep >= 24 h | L4 | docs.stripe.com |
| Quote/hold expiry + structured errors | expires_at on quotes and holds; 409 slot_unavailable for recovery | L4 | docs.stripe.com |
| Escalation path | requires_escalation -> continue_url to prefilled human checkout (UCP, Shopify) | L4 | shopify.dev |
| Signed agent identity (Web Bot Auth) | Verify RFC 9421 signatures and key directory; ChatGPT 'https://chatgpt.com', Google 'https://agent.bot.goog' (experimental); never trust UA alone | L3 | developers.cloudflare.com |
| Visa TAP / Mastercard Agent Pay | Agent signatures locked to merchant site and page; consumer identifiers; browse vs pay intent; built on Web Bot Auth | L5 | github.com |
| Payment mandates (AP2) | Signed Intent Mandate with price/timing rules lets agent create Cart Mandate when human not present; v0.2 Checkout/Payment mandates; non-repudiable proof for disputes | L5 | cloud.google.com |
| Delegated payment tokens (SPT / ACP) | Business receives a scoped token, never card details; ACP Delegated Payment spec for PSPs | L5 | docs.stripe.com |
| Strong customer authentication | Visa Payment Passkeys for agent-initiated payments | L5 | mynewsdesk.com |
| Audit logs | Log every tool call with correlation IDs; redact PII; log scope elevation events | L4 | developers.openai.com |
| Fraud checks | PSP fraud screening on agent payments; velocity limits per agent identity; flag new agents and high-value/short-lead bookings (our practice) | L5 | blog.cloudflare.com |
| Data minimisation | Tools return only what is strictly necessary; no internal IDs or secrets (ChatGPT rejection reason) | L4 | developers.openai.com |
| Prompt-injection hygiene | User-generated content treated as data; no instructions in tool output; annotations untrusted unless from trusted servers | L2 | modelcontextprotocol.io |
| Anti tool-poisoning | One official server URL listed in registry and site; version tool changes; plain factual descriptions | L4 | registry.modelcontextprotocol.io |
| Agent terms and policy page | Who may book/buy, limits, deposits, cancellation, liability, dispute contact (cf. Shopify Agentic Storefronts Supplemental Terms) | L3 | help.shopify.com |
The MCP specification's own security rules (must-haves for level 4): - no token passthrough: servers "MUST NOT accept any tokens that were not explicitly issued for the MCP server"; - per-client consent; - inputs validated, access controlled, calls rate-limited, outputs sanitised; - tool output is data, never instructions (prompt-injection hygiene); - one official server URL listed in the registry, so a fake copy can't poison the tools.
ChatGPT's app review checks read/write/destructive hints and data minimisation.
Payment safeguards come from the payment providers: - Stripe Shared Payment Tokens carry currency, max amount and expiry (≤180 days); - Visa has consumer-set limits and payment passkeys; - Google's AP2 uses signed intent and cart mandates.
Links are in the table above.
Same buyers for every level: where buyers already ask AI before buying, where firms already pay agencies, and where we can legally reach them. Software firms are also the first buyers for level 4 (Agent Door).
51% of software buyers now start research in an AI chatbot (G2 2026). IT gets the highest share of traffic from AI (2.80% vs 1.08% average). Two of your past clients are software products, so they can become the first case study.
51% of B2B software buyers start research in an AI chatbot (up from 29… · IT industry: 2.80% of site traffic from AI referrals vs 1.08% average;… · AI Overviews on 82% of B2B tech queries (Feb 2026, BrightEdge data via…
AEO Engine's core buyer: 15 of its 19 case studies are e-commerce. Retail AI traffic grew 393% a year (Adobe, Q1 2026) and converts better. One past client is a consumer brand.
15 of AEO Engine's 19 published case studies are e-commerce / consumer… · Adobe: retail AI-referred traffic +393% YoY in Q1 2026; AI visitors co… · Conductor: Consumer Staples 1.91% of traffic from AI referrals (2nd hi…
Exactly AEO Engine's own hook: its 4 monthly "Answer Indexes" all rank law firms. Only 7.4% of 216 UK solicitor firms studied were ever named by AI. Test with a copied "Family Law 100" index.
AEO Engine's 4 Answer Indexes are all law firms: 300 questions x 4 eng… · 2 of AEO Engine's 8 featured case studies are legal (e.g. My Rights La… · Legal AEO median $8,170/month, +42% vs 2025
Not a segment but the fastest start. 6 past clients plus one intro each; 136 Bali/Phuket villa managers with 20+ villas; iGaming B2B suppliers ("best crash-game provider" prompts are answered normally).
58% of travellers use AI tools to plan trips · Travel the most-cited AI category (~23%)
Copy AEO Engine's partner program: 20% of collected cash in year 1 and 10% in year 2 for referrals, or white label at 15% off. Agencies already sell SEO retainers to our buyers.
AEO Engine partner program: referral 20% of collected cash year 1, 10%… · White-label GEO/AEO wholesale: AgencyPlatform $178, 51Blocks $429 per …
Real AI demand, but smaller budgets (accountants, travel) or advertising rules that limit claims: no public Botox ads (CAP 12.12), FCA financial promotions.
Health Care: AIO on 48.75% of queries; 0.64% of traffic from AI referr… · Healthcare AIO coverage 88% (Dec 2025)
AI engines avoid or mishandle gambling queries (AI Overviews on ~1–12% of them), and regulators are watching. In one test, all 10 chatbots pointed users to unlicensed sites.
AI Overviews on ~1.4% (late 2025) to ~12% (Sep 2026) of gambling queri… · All 10 AI chatbots tested pointed consumers to unlicensed gambling sit…
First four bars: share of site traffic from AI referrals, scaled to 3%. Last two: share of buyers, scaled to 100%. Sources: Conductor AEO/GEO benchmarks, G2 buyer behaviour 2026, Omnisend via FashionNetwork.
Agency prices in these sectors (First Page Sage): B2B SaaS $8,640 a month median, legal $8,170, e-commerce $5,310, finance $9,850. Our $1,500 is a fraction of that, which is the copy's price edge.
How we count the pools. Companies House free bulk file (1 Oct 2026), active companies by industry code, "non-micro" = files more than micro accounts. Shopify stores from Store Leads (23 Oct 2026). Law LLPs carry no industry code, so they were counted by name.
Not yet measured: - how many UK software firms sell a product rather than dev work; - how many Shopify stores are limited companies with £1M+ sales.
Warm first, then a cold-email machine where every email carries the prospect's own AI snapshot. Legal only to companies (UK Ltd/LLP, Irish companies), never sole traders.
Personal note + their own snapshot + founding price in exchange for a consented case study. One named intro each (~1 more client from ~6 intros). Assumptions.
Cold email 200 UK Ltd web/SEO/PR agencies with AEO Engine's partner terms (20%/10% referral, white label −15%). 1 partner → 1–3 clients in 6 months (assumption).
Software firms and Shopify brands, 150 contacts each to start, then 1,050 a month. Keep the segment with the better reply rate after 300 contacts.
Copy AEO Engine's monthly ranking: "UK HR & payroll software 100", "UK skincare brands 100". Each firm's rank is the email hook ("you're #47 of 100").
Personal messages with snapshots; a Bali villa-manager index. No cold email to Indonesia/Thailand. Gambling-sector clients pay by wire only.
"You can email or text any corporate body (a company, Scottish partnership, limited liability partnership…)". Keep a do-not-email list.
Treated as individuals: only with consent. So many small Shopify stores and accountants are out: we filter to Ltd/LLP.
No hidden identity, a valid address to opt out (PECR reg. 23). Named work emails are personal data: legitimate-interest assessment, privacy line in email 1, instant suppression.
Allowed to a company's business address until it objects; identity shown; opt-out offered. Sole traders need opt-in.
No guaranteed #1, no unsupported "revenue opportunity", SRA rules for law firms, FCA rules for finance (see the pitch section).
Lists. - Software firms and law firms: the free Companies House bulk file (active; industry codes 62012, 58290, 63110 or 69102; non-micro). - Shopify brands: one month of Store Leads Pro (~$250), filtered to limited companies by the company number in the store footer. - Every email address is found and verified with Hunter or Instantly plus MillionVerifier.
Sending setup (from the round-7 kit): - 1 brand domain that never sends cold, plus 3 look-alike sending domains; - 6 Google Workspace inboxes with SPF, DKIM and DMARC; - 3–4 weeks of warm-up, ramping 10 → 25 emails a day per inbox; - Instantly ($97 a month); open tracking off. - Total ≈ $300–600 a month.
Pace. 1,050 contacts a month at published rates gives 0.4–1.7 new clients a month. 7 clients needs 4,375–18,065 contacts in total, so warm channels decide the first months. Doubling to 12 inboxes (2,100 a month) is allowed only while spam complaints stay under 0.1%.
The prospect sees their own numbers first, then the mechanism, then the plan. This is also how we explain the service on a call or in an email.
"AI repeats what the pages it reads say. We get you onto those pages, and we count the answers to prove it."
"Best {category} in {city}?" Half of buying questions in ChatGPT trigger a live web search.
We lock 150 such questions per clientIt fans out into several sub-searches in Google, Bing, OpenAI's and Perplexity's indexes, and Brave.
Bots allowed; pages that answer the sub-questionsMostly third-party pages: "best X" lists, reviews, press, YouTube, Reddit, plus your site.
We get you named on the pages it reads for your rivalsThe ones those pages agree on. It changes run to run, so one screenshot proves nothing.
We count share of answers across ~1,800 a week| Buyer question | ChatGPT | Perplexity | AI Overview | AI Mode | Gemini | Named instead |
|---|---|---|---|---|---|---|
| Best interior designer in Bristol? | ✗ | ✓ | ✗ | ✗ | ✗ | Studio Rival, Bristol Homes Co |
| Affordable interior design studio Bristol | ✗ | ✗ | ✗ | ✓ | ✗ | Studio Rival, DecorPlus |
| Who designs boutique hotel interiors in the South West? | ✗ | ✗ | ✗ | ✗ | ✗ | Studio Rival, West Design |
| Harbour & Pine vs Studio Rival | ✓ | ✓ | ✗ | ✓ | ✓ | — |
| How much does an interior designer cost in Bristol? | ✗ | ✗ | ✗ | ✗ | ✗ | Bristol Homes Co |
Illustrative layout with made-up names and numbers. A real snapshot is built from that company's own answers (~$0.45 of tracking fees each), shown as counts, not percentages.
We asked ChatGPT, Perplexity, Gemini and Google's AI 'best {category} in {city}' and 9 similar questions, twice each. {Company} was named in 6 of 100 answers. {Rival} was named in 41. Your one-page snapshot, no login: {link}
When AI recommends {Rival} for '{question}', it cites {site1}, {site2} and {site3}. {Company} isn't on any of them. Getting you onto pages like these is most of the job. Full list and what we'd pitch: {link}
Ask Perplexity '{branded question}' and it answers '{wrong fact}'. Your site says {right fact}. AI fills gaps with whatever page it finds, so one exact-facts page usually fixes it. What 5 AI tools say about you: {link}
Mostly, yes, and we say so; Google says AI features need 'no special optimizations'. The difference is the aim and the measure: AI runs several sub-searches and repeats the pages it reads, where third-party lists, press and reviews outweigh your own pages. We measure share of answers across 4 engines, which an SEO report doesn't show.
We lock the questions before starting, run 2 baseline weeks, and keep control questions we never work on. We report target lift minus control lift, with a dated log ('page X went live, Perplexity cited it on date Y').
It does. One screenshot proves nothing; under 1 in 100 repeat runs give the same list. So we count about 1,800 answers a week with a margin of error. Engine-wide changes hit the control questions too, so we can see them.
Keep them. We need read-only and draft-only access and don't touch their links or technical work. We add what they usually don't: AI answer tracking, getting you named on the pages AI cites, an exact-facts page, and AI lead attribution. We share our sub-query gaps with them.
Yes, because it's earned visibility: real pages, real reviews, labelled paid placements. We follow Google's spam policies (no cloaking, mass AI pages or parasite pages) and the UK DMCC and FTC review rules. You can't buy a place in ChatGPT's answers: its ads are labelled and run separately, so we don't sell that.
4 read-only or draft-only logins, a one-page fact sheet, about 1 hour in week 1, 15 minutes a month from one expert, and approvals within 3 business days. No passwords, no customer lists, no payment data.
Then we work free until the agreed number moves. You see the trend weekly, and we'll tell you at day 45 if it's flat. Answers from the model's memory lag, so we say upfront which questions are slow.
Ukraine sells about $6.7B a year of IT services abroad (2025, NBU). We invoice in USD or EUR, work in English, and are 2 hours ahead of London. All work lives in your Google Drive and CMS, so if we stopped you'd keep everything. By default we hold none of your customers' data, so there is no data-transfer issue; with GA4 Viewer access we sign a DPA plus SCCs/IDTA.
No one can, and AI rankings aren't stable anyway. We commit to share of answers on agreed questions.
Ads appear labelled and separate; they don't put you inside the answer. The two can run together.
Where the pitch comes from. AEO Engine's free report asks for a URL and work email, then shows: - a 0–100 score with a per-engine breakdown; - the top 3 rivals and share of voice; - a "revenue opportunity"; - the top 5 quick wins.
We copy the shape but drop the revenue-opportunity number unless we can back it (UK CAP code 3.7 requires evidence on file).
How we build the snapshot. 10 questions × 5 AI surfaces × 2 runs = 100 answers, about $0.45 of tracking fees. We show counts, not percentages, because 100 answers carries a ±6–10 point margin.
Other openers. A source-led one ("the 3 pages AI reads about your rival"), a facts-led one ("Perplexity thinks you cost X"), and one used when the prospect's robots.txt blocks ChatGPT's search bot (a one-line fix).
Four logins, all read-only or draft-only, a one-page fact sheet, about 1 hour of their time in week 1 and 15 minutes a month from one expert. Every item says which lever or which measurement it feeds.
| What | Why we need it | How we get it | Personal data | If they say no |
|---|---|---|---|---|
| Google Search Console | Their real search questions become our question set; the organic baseline; Google AI Overview/AI Mode clicks are counted inside these totals.feeds: measure: GSC, prompt set | ranking for sub-queries, crawler access | Restricted user: can view most data, can't change settings or submit sitemaps | no personal data | Build prompts from People Also Ask, third-party keyword data and templates; no organic baseline, so the traffic guarantee is off |
| Bing Webmaster Tools | The only free first-party report of AI citations: which of their pages Copilot and Bing's AI cited, for which questions.feeds: measure: Bing AI Performance | crawler access | Read-only user (can be limited to a URL path) | no personal data | Lose Copilot citation data; the client can export the AI Performance table monthly |
| GA4 (optional) | AI referral visits: chatgpt.com, perplexity.ai, gemini.google.com, copilot, claude.ai, and ChatGPT's utm_source tag.feeds: measure: GA4 AI referrals | Viewer with cost and revenue metrics hidden. The client's admin creates our "AI assistants" channel group (needs Editor). Or they share an aggregated Looker Studio report instead. | minimal personal data | Aggregated Looker/CSV report; if that is refused too, drop the referral metric and use GSC plus self-reported leads |
| What | Why we need it | How we get it | Personal data | If they say no |
|---|---|---|---|---|
| CMS draft role | We write answer pages, the facts page, comparison pages and refreshes in place; they publish.feeds: ranking for sub-questions, facts page, freshness | WordPress Contributor (can't publish or upload images) · Shopify collaborator limited to Blog posts and pages · Webflow Content editor with publishing off. Shopify and Webflow can't fully block publishing, so a contract rule covers it. | no personal data | Google Doc hand-off; the client pastes it (+1 day per batch) |
| What | Why we need it | How we get it | Personal data | If they say no |
|---|---|---|---|---|
| The 20–30 questions buyers ask | Combined with Search Console, People Also Ask and Bing grounding queries into the locked 150-question set.feeds: measure: prompt set | ranking for sub-queries | Typed list from sales and support, no transcripts or names | no personal data | Build from GSC, PAA and templates; fits the buyer less well |
| Exact brand facts | Name, founding, HQ, locations, products, prices or ranges. Becomes the facts page AI quotes, and our accuracy check on branded questions.feeds: brand facts page | measure: fact accuracy on branded prompts | Our fact-sheet template; the client approves | no personal data | We compile from the public site and company register; the client must still approve; accuracy scoring gets weaker |
| 3–5 rivals | Share of voice, comparison pages, and which rival's cited pages we target.feeds: measure: competitors, share of voice | earned third-party mentions, own comparison pages | Intake form plus rivals detected in the answers | no personal data | Use the rivals detected automatically from the answers |
| Markets, languages, priority products | Sets the country and language we ask the AI tools in.feeds: measure: prompt set | all | Intake form | no personal data | Default to HQ country, English |
| Banned claims and regulated limits | What we must never write (superlatives, health or financial promises, legal-services rules).feeds: compliance for brand facts page, stats/quotes, comparison pages, earned mentions | Intake form | no personal data | Strict default: no superlatives, no health or financial claims, no rival claims without evidence |
| One expert (15 min a month) | Journalists and list editors need a named person to quote. Without one, press work falls back to list outreach.feeds: earned third-party mentions, stats/quotes on page | Voice note or email; approves every quote in their name | minimal personal data | No journalist pitching in their name; PR falls back to list outreach; quotes attributed to the company |
| What | Why we need it | How we get it | Personal data | If they say no |
|---|---|---|---|---|
| Proof assets | Case studies, real numbers, awards, certifications: the facts that make pages and pitches citable. One new data point a month.feeds: stats/quotes on page, earned third-party mentions, brand facts page | A shared folder the client controls | minimal personal data | Lower volume; more scaled-content risk |
| Existing mentions (URLs only) | Past press, directories, lists; plus any paid placement, which we must disclose.feeds: earned third-party mentions | compliance | URL list | no personal data | We find them by web search; paid-placement disclosure becomes a contract warranty |
| Review-platform profiles | Which directories AI cites in their category and how they compare on reviews.feeds: reviews / earned third-party mentions | We read public pages | no personal data | Public audit only; no invite program |
| Sign-off on the locked questions + baseline | Fixes the question set and the 2 baseline rounds the guarantee is measured against.feeds: measure: prompt set, guarantee baseline | Signed baseline sheet | no personal data | No guarantee |
| What | Why we need it | How we get it | Personal data | If they say no |
|---|---|---|---|---|
| Let the AI search bots in | Sites that block OpenAI's search bot "will not be shown in ChatGPT search answers".feeds: crawler access | Their developer applies our robots.txt / CDN diff; we never take DNS or CDN admin | no personal data | Stop: we don't sign if the bots stay blocked |
| Submit sitemaps to Google and Bing + IndexNow key | Read-only roles can't submit sitemaps; faster discovery of new pages.feeds: crawler access | A 2-minute instruction to their owner or developer | no personal data | Slower discovery |
| "How did you hear about us?" with AI options | The only direct link from AI answers to leads: ChatGPT / Perplexity / Google AI / Gemini / Copilot / other AI.feeds: measure: lead attribution | They add the field to lead/booking forms; we receive monthly counts per option only, no names | no personal data | No lead attribution; report visibility and referrals only |
| Review invites to all customers | Real reviews on the directories AI cites. Never selective, never paid: UK DMCC Act and FTC rules.feeds: reviews / earned third-party mentions | Our template, their sending tool; they keep the customer list | no personal data | Lever 3 runs on lists and press only |
| What | Why we need it | How we get it | Personal data | If they say no |
|---|---|---|---|---|
| Public API spec or docs | We generate the MCP server's tools from it (one job per tool).feeds: level 4 tools | OpenAPI file or public docs URL | no personal data | We build read-only tools from public pages only |
| Sandbox key, read-only first | Test tools without touching real customers.feeds: level 4 tools | A sandbox / test-mode key with read scopes; write scopes later via OAuth with their login | no personal data | Public read tools only |
| Calendar / property-system export (services) | search_availability and get_quote read from it.feeds: level 4 booking tools | Read-only iCal or channel-manager export first; system API later | minimal personal data | Quote tool only, no availability |
| Limits and approver | The auto-accept rules (amount, nights, seats, per-day caps) and who approves above them.feeds: safeguards | Signed limits sheet + one named approver (WhatsApp or email) | minimal personal data | Every write goes to a human |
| Agent terms page | Who may book or buy via agents, deposits, cancellation, liability, dispute contact.feeds: safeguards, level 3–5 | We draft; they approve and publish | no personal data | No write tools go live |
Personal data: by default we hold none of their customers' data, so we aren't a processor. The one exception is GA4 Viewer: Viewers can open per-user reports. That needs a data processing agreement plus EU standard contractual clauses (or the UK IDTA), because Ukraine has no EU adequacy decision. So we offer an aggregated Looker Studio report first.
If the client refuses an item. - Most items degrade gracefully: Google Doc hand-off instead of CMS access, auto-detected rivals instead of a list, public review audit instead of invites. - Two items are hard stops: the AI search bots must be allowed, and the question set must be signed off before the guarantee applies.
Time asked of the client. About 1 hour in week 1 (logins, intake, fact sheet), 15 minutes a month from the expert, approvals within 3 business days.
Role docs: Search Console users · Bing users · GA4 roles · WordPress roles · Shopify permissions · Webflow roles · OpenAI bots
Level 1: count, don't screenshot. Headline = share of AI answers that name the client, on locked questions, against control questions. Levels 2–4: agent task tests run by hand on the client's own site, plus automated checks we're allowed to run.
Headline: AI names the client in X% of ~7,200 answers a month (±3–4 points), minus the change on control questions = our effect.
answers naming the client ÷ all answers (per engine, then averaged)AI names them more often. The headline, always with ± margin and the control-question change.
client mentions ÷ (client + 3–5 rivals' mentions)They gain on rivals, not just ride a market-wide rise.
answers citing a client URL ÷ all answers; answers citing a page WE got them ontoThe strongest link to our work: AI uses pages we created or placements we earned (dated in the action log).
correct key facts ÷ facts stated on branded questionsAI says true things about them (vs the facts page).
tasks completed by agents ÷ tasks attempted; steps per task; blockersA buyer's agent can actually get the price, start the trial or hold the booking, not just read about the business.
tool calls by agent identity; quotes → holds → bookings; escalations; errorsAgents are using the API, and how often it turns into business.
Search Console Generative-AI report impressions; Bing AI Performance citationsFirst-party proof from Google and Microsoft that their AI shows the client's pages (impressions only, no clicks).
sessions in GA4's AI Assistant channel + our regex; key events ÷ sessions vs organicAI sends visits that convert. A floor: apps often strip the referrer.
leads whose "How did you hear about us?" mentions an AI ÷ all leadsAI-driven demand analytics misses.
won deals with AI in self-report or AI channel × deal valueMoney. Never promised.
| How we collect AI answers | Same as real users see? | Cost per 1,000 answers | Allowed? | We use it for |
|---|---|---|---|---|
| Ask the AI apps ourselves, by hand | exactly | ~$500 of VA time | normal use | Monthly proof pack (~30 screenshots per client), calibrating the tracker, sales snapshots. Too slow for the headline (±10–15 points). |
| Script the apps ourselves (browser automation) | exactly | ~$0 | forbidden | Never. OpenAI, Google and Perplexity terms forbid automated extraction from their apps. |
| Tracker vendor reading the apps (DataForSEO) our stack | close | $1.2–4 | vendor's risk | Main weekly numbers for ChatGPT, Gemini, AI Overviews, AI Mode. The vendor does the reading under its own terms; our risk is the feed being cut. We disclose the method in reports. |
| Official APIs with web search (Perplexity, OpenAI, Gemini) our stack | differs | $5–30 | allowed | Perplexity (its API matches its app well), backup if the vendor fails, a quarterly "from memory" test. ChatGPT/Gemini APIs overlap with app answers on only 15–32% of brands. |
| Bing Webmaster Tools AI Performance our stack | first-party | $0 | allowed | Copilot citations; its grounding queries become new question ideas. |
| GA4, Search Console, the client's CRM our stack | first-party | $0 | allowed | AI visits, Google AI impressions, brand searches, leads that say "ChatGPT". |
margin = 1.96·√(p(1−p)/n_eff), where n_eff discounts repeat runs and engines that move together (assumed correlations 0.5 and 0.2; recomputed from the first two baselines). More questions shrink the margin far more than more runs. The control-adjusted effect has a wider margin (≈7–9 points with 40 control questions).
This month AI assistants named Example Ltd in 18% of ~7,200 answers to buyer questions (±3.4 points), up from 9% when we started. Control questions rose 2 points, so about 7 points are our work (not yet certain at this sample size). 212 answers cite pages we got them onto. ChatGPT and Perplexity moved; Google AI Overviews didn't. Next month: 3 actions. What didn't work: 1 line.
Illustrative numbers. Every report also shows: a weekly line (target vs control vs rivals, 95% band, action dates, shaded model-change weeks), per-engine before/after, a question × engine heatmap, the gap list of sites AI cites for rivals but not the client, the method box, the screenshot proof pack and the raw answers.
How we show "it was us": - the questions, rivals, controls, market and method are locked before the baseline; - a dated action log; - target lift minus control lift, claimed only when its margin excludes zero; - cited URLs sorted into "ours", "our placement" and "other"; - lift that appears after the action date, not before; - model-change weeks shaded.
What we can't claim: - a rank or "#1"; - what logged-in users see; - that uncited mentions came from us; - clicks from Google AI Overviews or AI Mode (Google doesn't separate them); - revenue caused by AI.
Known measurement traps: - AI apps often strip the referrer, so GA4 AI traffic is a floor. - ChatGPT's tracking tag broke on 8–15 Oct 2026; every report notes it. - Search Console's Generative AI report shows impressions only. - Bing shows citations, no clicks.
Cost per client per month: ≈ $47–55 at 150 questions (DataForSEO ~$12, Perplexity API ~$10, parsing, hosting ~$10, VA hand-check pack $12–19), or $31–39 at 50 questions. The Anthropic API is never used; Claude is spot-checked by hand only, on the subscription.
Not yet confirmed: the wording of the OpenAI and Perplexity terms came from search snippets. Re-read them before putting it in a contract.
Agent-readiness score (levels 2–4). Each month, by hand, on the client's own site with consent: - the same 3 tasks (e.g. "20 seats annual in GBP, start a trial") run through ChatGPT agent, Claude in Chrome and Perplexity Comet; - for each, we record success, steps and the blocker.
Automated checks we're allowed to run: - robots and CDN bot-rule audit, and whether signed ChatGPT requests pass; - structured-data validation (page price = markup = feed); - raw-HTML vs rendered diff; - Chrome Lighthouse's experimental Agentic Browsing category; - form fillability; - for level 4: MCP conformance tests and the MCP Inspector.
Not allowed: scripted bulk runs of consumer agents, and tests on competitors' sites or CAPTCHA bypass.
Sources: Lighthouse agentic browsing · MCP conformance · Cloudflare AI Crawl Control
From signature to the day-90 review. AI does the bars marked AI; you sign, approve and talk.
The guarantee clock. It starts after the 2 baseline rounds, which need the signed question set and the bots allowed. A site migration voids it. If the number hasn't moved at day 90, month 4 is free. Working a month free costs only ~$170–470 plus hours.
When results show up. Search-based answers can move in days to weeks once a cited page names the client. Answers from the model's memory move only with model updates, which take months, so the report labels which questions are slow.
Every link on this page, re-checked when it was built.
Evidence labels. ICO, legislation.gov.uk, Google, Microsoft and OpenAI pages are primary. Conductor, Adobe, G2, BrightEdge, Similarweb, First Page Sage and Store Leads are vendor data. AEO Engine, Peec and Profound numbers are their own claims. Pools are our counts from the Companies House bulk file (1 Oct 2026).
Assumptions shown as such: the reply → paid rate (12.5–25%), warm-channel outcomes, and the correlations behind the margins (recomputed from the first two baseline rounds).
Example layouts. The snapshot and the monthly report on this page use made-up names and numbers to show the format.