Short answer. Two numbers decide the argument, and they point in opposite directions. AI assistants combined still send only a small single-digit-percent slice of search referrals, and users click a cited source inside an AI Overview only about 1% of the time it appears. But the referrals that do arrive convert well above organic search, according to referral-traffic studies from 2026. Both facts are true. A business case built on either one alone is dishonest, and the honest case is not a traffic case at all.
Most material on answer engine optimisation opens with a growth chart. The honest place to open is the referral share, because it is the number a finance function will find on its own, and finding it after the budget is approved is how a programme gets cancelled in its second quarter.
This piece sets out both sides of the ledger, then builds the version of the case that survives being checked. Answer engine optimisation (AEO) is the practice of structuring content so AI assistants can retrieve, understand and cite it accurately. Generative engine optimisation (GEO) is the closely related discipline of earning inclusion in the generated answer itself, across the sources an assistant draws on beyond your own site. If you want the fuller definitions and how the two relate, see what generative engine optimization is.
Key takeaways
- AI assistants send a small fraction of overall search referral traffic today; Google remains the dominant referral source by a wide margin.
- Users click a cited source inside an AI Overview roughly 1% of the time it appears, and overall click-through drops by roughly half on queries where an AI summary shows up.
- Referral traffic that does arrive from AI assistants converts noticeably higher than average organic search traffic, though the size of that premium varies by assistant and by study.
- The defensible AEO business case is a description-and-citation-rate case, not a traffic-growth case: most of the value is in how a buyer who never clicks is told about you.
- Roughly 85% of AI answer citations point to third-party sources rather than a brand's own site, which is why third-party presence is usually the largest line item in the budget.
Why is the traffic number not the whole case?
Three things push the other way, and they are the reason serious companies are still funding the work.
The visits are qualified to an unusual degree. Goodie's 2026 AI Search Traffic Report measured referral conversion by assistant and found conversion rates on ChatGPT and Perplexity referrals running well above the organic-search average, with Claude referrals showing the highest average session value of the assistants tracked. The relative ranking between assistants shifts between measurement periods, so treat any single number as a directional signal rather than a fixed rate — but a multi-fold difference means a small stream of AI referrals is not proportionally small in pipeline. Programmes that track this properly, the way share of answer is measured, watch the rate move quarter to quarter rather than quoting one snapshot forever.
The click was already leaving. Zero-click search already accounts for a majority of queries, and Pew's study of roughly 68,000 real search queries found users clicked a traditional result on 8% of searches where an AI summary appeared against 15% where none did — a drop of roughly half. The counterfactual for many of these queries is not a click to your site. It is no click to anyone.
The description is the product. When the buyer does not click, what they take away is the sentence the engine wrote about you. Being absent means being described by whoever else was cited, which is usually a competitor or a review aggregator — a dynamic covered in more depth in where AI citations actually go. That exposure exists whether or not you fund anything, which is what makes doing nothing a position rather than a saving.
How do you build a number that survives a CFO?
Do not model AEO as traffic. Model it as influenced pipeline, and be explicit about every assumption — including the ones you cannot measure.
| Input | Where it comes from | The trap |
|---|---|---|
| Buyer questions per month in your category | Your own question set, sized against search volume for the same intents | Using total keyword volume as if all of it triggers an AI answer |
| Share of those triggering an AI answer | Measured on your own list | Quoting a published average drawn from a different query mix |
| Your citation rate across repeated runs | Your own measurement, per engine | Reporting a one-off check as if it were a rate |
| Value of being described accurately | Modelled, and labelled as modelled | Presenting it as attributable revenue |
| Referral clicks and their conversion rate | Analytics, with AI referrers segmented | Assuming last-click captures the influence |
The formulation that holds up in a budget meeting sounds like this: in our category, a defined number of buyer questions a month are answered by an assistant; we are named in a measured share of those runs today; every one of those answers is read by a buyer whether or not they click; the referrals we do get convert meaningfully above organic; we are funding this to control how we are described, and tracking clicks as a secondary benefit. That approach, and the metrics that support it, are set out further in measuring GEO success beyond clicks.
That sentence survives scrutiny. "AEO will grow traffic 30%" does not.
Why is the number expected to move?
The current referral share is small. The trajectory is the argument, and it should be presented as a trajectory rather than as a fact about today.
- Gartner forecast in 2024 that traditional search engine volume would fall roughly 25% by 2026, while noting the figure reflects scenario modelling rather than a certainty.
- The category is fragmenting. Similarweb's generative-AI traffic tracking shows ChatGPT's dominant early lead giving up meaningful share to Gemini and Claude over the past year. A programme scoped to one assistant is measuring a shrinking slice.
- Click-through has not collapsed monotonically. Seer Interactive's tracking, compiled by Omnibound, found click-through on AI Overview queries fell sharply through late 2025 before partly recovering into 2026, still running below click-through on queries with no AI Overview.
Read that third point carefully, because it cuts against the simple narrative in both directions. Anyone quoting only the fall is selling urgency. Anyone quoting only the recovery is selling complacency.
What does the work actually cost?
- Measurement. A defensible programme runs a fixed question set repeatedly across engines and markets. Entry-level tooling in this market is inexpensive relative to the labour; the larger line by a wide margin is designing and reading the question set. This is the same discipline behind how to measure AI visibility without fooling yourself.
- Content revision. Usually cheaper than new content, and better supported by evidence. Recency is a strong retrieval signal, and the pages that already answer buyer questions are the ones worth revising first.
- Third-party presence. The largest and least controllable cost, and the one addressing the roughly 85% of AI references that point away from your own domain.
- Crawler access. Effectively free, and frequently the entire problem. A site that retrieval crawlers cannot fetch cannot be cited at any budget.
Note the shape of that list. The cheapest item is often the binding constraint, and the most expensive item is the one nobody controls. A proposal whose cost is dominated by a tool subscription has inverted both — a symptom covered in what GEO services actually cost.
What does the honest case not claim?
It does not promise traffic, placement, or stability, because none of those are things a vendor controls.
- It does not claim attributable revenue. AI referrals are small, frequently arrive without a referrer, and the influence happens in a session you never observe.
- It does not claim a guaranteed citation. No engine sells placement, accepts a submission, or offers an index request. Anyone promising one is promising something they do not control.
- It does not claim stability. A large share of an assistant's cited sources change from one run to the next, so the deliverable is a rate, not a fixed position.
- It does not claim this replaces search. Google still sends the large majority of search referrals. AEO is an addition to a functioning search programme, not a substitute for one.
How does Lifewood approach this work?
Lifewood scopes this work as a description programme with a measured rate attached, not as a traffic forecast, and says so in proposals.
The instrument is in-house: a fixed set of buyer questions per market, run repeatedly rather than checked once, reported per engine with retrieval and memory surfaces kept apart and the raw answers retained. Two figures drive the shape of the plan rather than the pitch. Because roughly 85% of AI references point at third-party sources, the on-site workstream is scoped as the smaller half from the outset. And because the retrieval surface responds to publishing in weeks while the memory surface changes only when a model is retrained, the two are budgeted on different horizons.
Multi-market programmes are where the cost model diverges most, because a question set has to be written natively in each market rather than translated. 100+ languages, 40+ delivery centres across 30+ countries and 56,000+ registered contributors are what make in-market authorship a staffing decision rather than a translation line. See AEO services and GEO services.