How Do You Forecast SEO Traffic When AI Answers Take the Clicks?

Can You Still Forecast SEO Traffic in 2026?

Yes, but not with the click curves most teams still use. The old model assumed a ranking position mapped to a predictable click rate. AI answers broke that link. A forecast built on 2023 assumptions will overstate traffic on exactly the queries where AI answers appear most often.

We get asked for traffic projections on almost every site build. Somebody needs a number for a board deck, and the honest answer is more useful than a confident one. The good news is that a defensible forecast is still possible. It just needs a different shape.

Here is how we build one, what the public data supports, and where we tell clients the model stops being trustworthy.

What Do the Click Numbers Actually Show?

Two independent studies point the same way. The Pew Research Center tracked browsing from 900 US adults in March 2025 and found that when a Google search produced an AI summary, users clicked a traditional search result in 8 percent of visits. Without an AI summary, they clicked in 15 percent.

Pew's sample was substantial. The study covered 68,879 unique Google searches, of which 12,593 produced an AI summary. Around one in five searches produced one, and 58 percent of respondents saw at least one during the month. Pew published the findings on 22 July 2025.

Ahrefs found a similar pattern from the publisher side. In research published on 4 February 2026, it compared 300,000 keywords, half with an AI Overview present and half without, using aggregated Google Search Console data. It reported that the presence of an AI Overview correlates with a 58 percent lower average clickthrough rate for the top ranking page.

Why Do Old CTR Curves Break Your Forecast?

Because the curve moved, and it moved unevenly. Ahrefs reported that average position one clickthrough rate for informational keywords without an AI Overview fell from 0.076 in December 2023 to 0.039 in December 2025. For AI Overview keywords it fell from 0.073 to 0.016 over the same period.

Read those four numbers slowly. Position one CTR dropped on both sets. It dropped roughly by half where there was no AI Overview, and by roughly three quarters where there was one. A single blended curve cannot represent both.

This is why so many forecasts missed in the last two years. They were not wrong about rankings. They were right about rankings and wrong about what a ranking is worth. If your model still multiplies search volume by a position one rate near 0.30, it is describing a web that no longer exists.

How Do You Split Keywords by AI Overview Presence?

Start by tagging every keyword in your target set with whether an AI Overview currently appears. Most rank tracking tools expose this as a SERP feature flag. Export the list, add a simple yes or no column, and treat the two groups as separate forecasts from that point on.

Then look at query shape, because it predicts the flag. Pew found that 60 percent of question format searches produced an AI summary, and 53 percent of searches with ten or more words did. Long, conversational, question shaped queries are the most likely to trigger an answer box.

This has an awkward implication for content plans built around question keywords. Those are precisely the queries most exposed to click loss. It does not mean stop writing them. It means do not model them at the same click rate as a commercial or branded query.

What CTR Should You Model For?

Use your own data first, and public benchmarks only as a sanity check. Pull the last twelve months from Google Search Console, split by query, and calculate your real clickthrough rate at each position band. Your site, in your niche, is a better predictor than any published average.

Where you have no history, the published figures give you a defensible floor and ceiling. A cautious model for an AI Overview keyword at position one uses something close to the Ahrefs figure of 0.016. A non AI Overview informational keyword sits nearer 0.039. Both are far below what planning templates usually assume.

Always model a range rather than a point. We present three scenarios: a low case using the AI Overview rate across the whole set, a base case using the actual split, and a high case using the non AI Overview rate. The spread itself is the honest part of the answer.

How Do You Handle Impressions That Never Become Clicks?

Count them separately and say what they are worth. Pew found that clicks on links inside the AI summary itself happened in just 1 percent of visits. So appearing in an AI answer usually produces a mention rather than a session. That is brand exposure, not traffic, and mixing the two ruins a forecast.

Google's own tooling reinforces the split. The Generative AI performance report in Search Console reports impressions only. Google's documentation states there are no clicks, no clickthrough rate and no position data in that report. You cannot convert those impressions into a traffic number, and you should not try.

Pew also measured what happens after. Twenty six percent of visits to pages with an AI summary ended the browsing session, against 16 percent without. The search is being satisfied on the results page. Your forecast should treat that as demand you can influence but not capture.

What Should the Forecast Actually Promise?

Sessions from non AI Overview queries, with a stated range. Everything else belongs in a second column labelled visibility, measured in impressions and mentions. Two columns, two units, no blending. The moment you blend them you have produced a number that cannot be checked later.

This structure survives contact with a finance team. They understand a conservative revenue line with a separate awareness line. What they will not forgive is a single traffic figure that misses by half and cannot be explained.

It also changes the conversation about content. Once the low click queries are visibly separated, it becomes obvious which pages exist to be cited and which exist to be visited. Those are different jobs and they deserve different success metrics.

How Do You Tell a Forecast Miss From a Traffic Drop?

Look at impressions before you look at clicks. If impressions are flat or rising while clicks fall, the market changed and your forecast assumption was wrong. If impressions fell too, something happened to your rankings or your indexing, and that is a different problem entirely.

Ahrefs described this gap in its data, and we see it constantly in Search Console. A page can hold its position, keep its impressions and still lose most of its sessions, purely because an answer box appeared above it. No technical audit will find a fault, because there is no fault.

Running that check first saves weeks. We walk through the full diagnostic sequence in our guide to diagnosing an SEO traffic drop, and the first branch is always this one.

What Does a Defensible Forecast Look Like?

Six inputs, all traceable. Your keyword set with an AI Overview flag on each. Current position for each. Your own clickthrough rate by position band from Search Console. A stated assumption for how positions will move. A stated source for every rate used. A low, base and high scenario.

Write the assumptions on the same page as the numbers. Not in an appendix. When the forecast is reviewed in six months, the assumptions are the only part anyone can usefully argue about, and burying them makes the review pointless.

Refresh the AI Overview flags quarterly at minimum. Coverage of that feature has kept moving, and a flag set that is a year old is closer to fiction than data. If you need a refresher on pulling the underlying numbers cleanly, our Google Search Console guide covers the exports we rely on.

How Should You Present This to Your Board?

Lead with the range, not the midpoint. Show the two columns, explain that one is sessions and one is visibility, and name the source behind every rate on the slide. A board that can see where a number came from will forgive it for being conservative.

Then set the review date in the same meeting. A forecast without a scheduled review is a promise, and promises about search traffic in 2026 age badly. A quarterly check against actuals turns the same document into a working model.

The strategic point underneath all this is that fewer clicks per impression is now the baseline condition, not a temporary dip. Planning around it is more useful than hoping it reverses, which is the argument we make in our piece on building a zero click search strategy.

If you want a second pair of eyes on a forecast before it goes in front of your board, we are happy to look at the model with you. Find us at phoenix.studio.