Direct answer: do not build one generic “AI traffic” landing page or send every referral into a chatbot. Build a short, observable route from the page that was actually cited to the likely language and market, the job the visitor needs to complete, the required evidence, an appropriate next step, and the later CRM outcome. Preserve source, landing page, and consent; ask for missing context instead of reconstructing the prompt; keep visit, contact, qualification, opportunity, and revenue as separate evidence grades. AI may classify stated context and recommend a route, but it should not invent intent or receive automatic credit for the deal.
The gap appears after a GEO win. An AI answer may cite a technical article, the corporate about page, or the homepage. The visitor arrives with a narrow question and meets broad navigation, the wrong locale, proof for another market, a generic form, or a meeting CTA that asks for too much too soon. Publishing more sources may improve discovery but does not repair this step. https://makinai.co/insights/en/company-search-absent-ai-answers-geo-plan covers discovery, and https://makinai.co/insights/en/measure-ai-search-visibility-pipeline-geo separates presence, referral, and pipeline. This article scopes the conversion route between them.
Write an AI Referral Route Contract
- Observable source — referrer, UTM, landing page, timestamp, and available consent; never infer the full prompt from these signals.
- Entry context — topic, entity, offer, and evidence the page really contains, including version and market.
- Language and market — declared preference, URL, browser, and country only as signals; offer a visible switch and preserve the choice.
- Visitor job — compare, verify, estimate, diagnose, implement, or talk; ask when ambiguous.
- Evidence and eligibility — proof, limitations, region, industry, size, integration, and constraints that determine fit.
- Next step — read a source, compare a service, prepare a brief, request a diagnosis, or open contact; each step asks only what it needs.
- CRM handoff — original source, latest source, landing page, content, language, market, stated need, and permission in separate fields.
- Outcome — useful action, confirmed contact, qualified lead, opportunity, revenue, and cost with a window and owner.
The contract prevents two shortcuts: treating chatgpt.com as one intent, and overwriting every object with “AI” as its source. OpenAI says ChatGPT adds utm_source=chatgpt.com to referral URLs. GA4 can combine landing page with session source/medium. That supports session observation. It does not reveal the complete prior conversation, prove the citation caused a purchase, or justify filling industry, title, and urgency the visitor never stated.
Match investment to the job
- Level 1, editorial route — map entry pages to one useful next-action block, local links, a language alternative, and CTA events. Best for low volume and a simple offer.
- Level 2, operational route — preserve parameters, offer market choice, use progressive forms, record generate_lead, and carry context into CRM. Best when marketing and sales lose evidence at handoff.
- Level 3, adaptive route — classify a stated need, recommend a service, or prepare a brief from approved content with fallback. It requires volume, taxonomy, testing, ownership, and mature outcomes.
Level 1 may need no model. A page→job→proof→CTA table can solve the first bottleneck. At Level 2, transparent rules usually handle language, served geography, service, and consent. Use AI at Level 3 when free text and portfolio breadth create real variability; require reasons, correction, and abstention. https://makinai.co/insights/en/ai-agents-read-website-webmcp-when-let-act explains when a site tool may prepare an action without submitting it autonomously.
Example: from a cited article to a useful brief
Imagine a U.S. B2B software company with a cited guide about CRM migration. A visitor lands on the technical article from an AI answer. The page offers three next steps tied to the job: assess data readiness, compare architecture choices, or discuss implementation. If the visitor chooses implementation, a progressive form carries the landing page, topic, observed source, and selected task; it asks for the current system, objective, timing, and permission. CRM preserves original and latest source separately. Sales receives declared context, not a guessed prompt. This is a hypothetical workflow, not a client case or measured result.
For U.S. buyers, meeting booking may be appropriate only after eligibility and owner capacity are known. A fast calendar link that ignores state, product, account ownership, or sales coverage can create abandonment and internal rejection. Applicable privacy and recording rules vary; this is not legal advice.
Pilot five pages in six weeks
- Week 1 — select five commercially relevant or referred landing pages, preserve a baseline, and list the questions each truly answers.
- Week 2 — audit locale, hreflang, x-default, canonical, CTA, proof, market, consent, events, and CRM fields.
- Week 3 — write the route contract, page→job→next-step map, and fallback rules; remove unobservable inferences.
- Week 4 — implement next-action blocks, locale choice, parameters, events, and fields; test mobile, blocked cookies, direct entry, and returning users.
- Week 5 — expose a limited share, review wrong routes, abandonment, duplication, and CRM handoff; retain a stable version.
- Week 6 — inspect useful action, confirmed lead, quality, speed to service, and cost; keep, repair, expand, or remove.
Deliverables include the entry inventory, route contract, localized map, content and CTAs, event dictionary, CRM fields, test cases, dashboard, runbook, and backlog. Google recommends reciprocal hreflang and x-default, but annotations cannot choose the right offer or repair a page without market-relevant proof. Google also documents generate_lead; fire it only when the business has defined what a lead is, not for every click.
Measure the loss between referral and outcome
Track observable referred sessions; landing pages; displayed and chosen locale; language switches; read, compare, or prepare actions; CTA starts; form opens; confirmed contacts; generate_lead; qualification; opportunity; mature revenue; time; and cost. Guardrails include wrong redirects, inferred prompts, invented fields, overwritten sources, duplicates, opt-outs, unsupported CTAs, partial submissions, context-free handoffs, and duplicated attribution.
MAKINAI's aggregate snapshot for October 11 through November 7, completed November 8, recorded 194 sessions, 10 organic clicks, 2,882 impressions, and zero lead events. Three identifiable sessions came from ChatGPT: one to the homepage, one to the about page, and one to an article; two were engaged and none produced a lead. This is a small single-site observation affected by tracking blocks, missing referrers, and Google reporting delay. It does not prove the route caused a loss or support an “AI traffic” conversion rate. It only justifies inspection.
HubSpot now distinguishes Original and Latest Traffic Source and documents AI Referrals, while warning that blocked cookies affect values. Keep record source, declared origin, and influence separate too. When a buyer says an AI answer introduced the company, retain that declaration without replacing technical source. If the deal matures months later, the evidence may show contribution, not causality.
Cost grows with pages, locales, markets, services, CTA variants, login state, consent, event quality, forms, CRM integration, deduplication, testing, volume, and ongoing operation. Content and CRO own the page; SEO/GEO owns entry; analytics owns observation; RevOps preserves context; sales validates quality; engineering implements; privacy reviews purpose. The agency should own the complete route or make every handoff explicit.
If AI referrals reach broad pages and disappear before a useful conversation, bring MAKINAI at https://makinai.co/services/en/data-content-intelligence-systems five landing pages, current CTAs, the event dictionary, and CRM source fields. We can scope the smallest measurable route. For the broader conversion system, use https://makinai.co/insights/en/ai-improve-website-conversion-without-replatforming. The goal is not to give AI credit; it is to keep a visitor with a specific question from meeting a website with no appropriate next step.