Forecasting the channel switch from search to AI

The forecast problem

Every client wants the same number for 2027: how much search traffic moves to AI, and what it is worth. The tools they use to answer it were built for a different world.

A search forecast rests on two inputs. Last year's sessions by channel, and Google Keyword Planner for demand. Both assume the click still exists. Neither knows a prompt from a query. Keyword Planner has no AI equivalent and Google will not build one.

So the forecast needs a different shape. Search metrics still matter because clients understand them. The job is to map each one to its AI counterpart, put a ratio on the link, and test that ratio against real data. This article sets out that model. Every number in it is a starting position, chosen to be argued with.

Why crawl counts fail as a session proxy

The obvious shortcut is to count AI bot hits in the logs and treat growth in crawling as growth in demand. It fails on its own. Most AI crawling is for training, and training returns nothing. Cloudflare's Q1 2026 data put only 8% of AI crawler traffic on search and 2.2% on live user queries. The rest builds models.

The ratios also move too fast to bank. Cloudflare Radar had Anthropic at 3,386 crawls per referral in June 2026 and 1,917 in July. OpenAI went from 647 to 251 in the same month. A crawl forecast breaks the day a platform ships a search feature.

Split the crawlers by purpose and the idea works. Cloudflare now classifies each crawler as Training, Search or Agent. Server logs give the same split by user agent. The fetches that matter are the user-triggered ones: ChatGPT-User, Perplexity-User, Claude-User. Each fires because a real prompt retrieved your page. That is the AI equivalent of an impression. Count those, ignore the rest.

The crawl-to-refer ratio keeps one job. It is a leading indicator of platform behaviour. When a platform's ratio collapses, its referral curve is about to bend.

The ladder

Every search metric has an AI counterpart. Put them side by side and the forecast has a structure a client already understands.

Search metric	AI counterpart	Where it comes from
Search volume	Prompt volume	Search volume for the intent bucket, times the share of that intent moving to AI
Impressions	Presence rate	Share of tracked prompts where the brand appears, from an AI visibility tool
Ranking	Share of voice and citation position	Same tracked prompt set
Clicks	AI referrals	GA4 referrals from chatgpt.com, perplexity.ai, gemini and copilot, plus user-agent fetches in the logs
CTR	Citation-to-referral rate	Referrals divided by citations
Conversion rate	AI referral conversion rate	GA4, by channel

Two rungs need stating plainly. There is no Keyword Planner for prompts, so prompt volume is derived. And the citation-to-referral rate is the number clients get wrong. Most AI answers end the journey. A citation is a mention with a link, and most of those links go unclicked.

The straw man

Seven rungs. Each has a starting ratio, the evidence behind it, and the test that proves or kills it. The ratios are deliberately round. Their job is to be replaced by your own data.

Rung	Straw man	Basis	Test
1. Search volume to prompt volume	20% overall: 30% informational, 15% commercial, 5% transactional	AI Overviews trigger on 41% of informational and 34% of commercial queries; 67.8% of consumers used AI for product research in the last 30 days	Year-on-year GSC impression change per intent bucket, split by whether an AI Overview shows
2. Prompt volume to presence rate	25% for a category leader, 10% for a challenger	Top 10 brands take 47% of citations in their industry; the long tail gets 12%	100 tracked prompts a month in the visibility tool, presence rate by platform
3. Presence to linked citation	40%	Perplexity cites in 13.8% of answers, ChatGPT in 0.7%. Blended by traffic share the linked rate is low, so 40% of appearances is generous	Citations divided by appearances in the same tracked set
4. Citation to referral	7%	93% of AI search sessions end without a click	GA4 AI referral sessions divided by monthly citations; cross-check against user-agent fetches in the logs
5. AI referrals as share of organic	2% now, 3x growth next year, 2x the year after	1.84% of organic traffic in March 2026, up 8.7x in 24 months; 1.08% of all traffic with 300% annual growth	Fit a logistic curve to 12 months of your own GA4 AI referral sessions
6. Organic CTR displacement	Minus 30% informational, minus 15% commercial, minus 5% transactional	25.11% of Google searches now carry an AI Overview, up from 13.14% in March 2025	GSC CTR change on queries with an AI Overview versus matched queries without
7. Conversion multiplier	2x organic, range 1.3x to 2.5x	4.21% AI referral conversion against 1.94% organic; Adobe shows AI shoppers converting 42% better in March 2026; Contentsquare sits lower at 1.3%	GA4 conversion rate by channel, AI referrals versus organic

Rungs 1 to 4 describe the funnel from demand to a visit. Rung 5 is the shortcut: skip the funnel and fit a curve to referrals you can already see. Rungs 6 and 7 turn visits into money. A client with 12 months of GA4 data can test rungs 4 to 7 this week. Rungs 1 to 3 need a visibility tool and a tracked prompt set.

A worked example

Take a site with 1,000,000 organic sessions a year, split 60% informational, 30% commercial, 10% transactional, converting at 2%.

Line	2026	2027 straw man
Organic sessions	1,000,000	820,000
Informational lost to displacement (rung 6, minus 30% of 600,000)		minus 180,000
Commercial lost (minus 15% of 300,000)		minus 45,000
Transactional lost (minus 5% of 100,000)		minus 5,000
AI referral sessions (rung 5, 2% then 3x)	20,000	60,000
Organic conversions at 2%	20,000	16,400
AI conversions at 4% (rung 7)	800	2,400
Total conversions	20,800	18,800

The honest reading is uncomfortable. On these ratios the site loses 230,000 organic sessions and gains 40,000 AI sessions. Conversions fall 10%. The 2x conversion multiplier softens the blow. It does not cancel it.

That is why the ratios must be tested rather than believed. Two of them move the answer most. Rung 6, the displacement rate, sets the size of the hole. Rung 5, the growth curve, sets how fast AI fills it. A site whose AI referrals grow 5x rather than 3x, and whose informational CTR falls 20% rather than 30%, breaks even. Both are measurable from the client's own GSC and GA4 data today.

How to run it

The forecast takes a week with the right access. Six steps.

Baseline last year's organic and paid sessions by intent bucket. Classify the top 1,000 queries by hand if you must. Intent is the axis everything else hangs on.
Measure your own displacement. Pull GSC CTR for queries with an AI Overview and matched queries without. The gap is rung 6. Use it in place of the straw man.
Fit a growth curve to AI referrals. Build a GA4 channel group for chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com and claude.ai. Twelve months of data is enough. Fit logistic, not linear. Adoption saturates.
Measure the conversion multiplier from the same channel group. Use the client's number, not the industry one.
Add the purpose-split crawl data from the logs as a leading indicator. Rising user-agent fetches predict rising referrals a quarter out.
Publish three scenarios. Low, central, high. Show cannibalised sessions and incremental sessions on separate lines. A point forecast in this channel is fiction and the client knows it.

Rungs 1 to 3 improve the model over time. They need a tracked prompt set and a visibility tool, and the numbers only mean something after three months of history. Start them now so next year's forecast has them.

Test it in public

The straw man will be wrong. That is the point. Every ratio above is a claim you can check against a real site in an afternoon. I will publish the results as clients let me. If your numbers disagree, send them. The model gets better each time it loses an argument.