AI Mode cuts clicks and Search Console hides it, so measure outcomes

If a query triggers an AI answer in Google, its average position in Search Console no longer tells you much. People pushed into Google's AI Mode click through to external websites less. Search Console also gives every link inside an AI Overview the same position, and because the box usually sits at the top of the page, that position is usually 1.

Those two facts together turn the most familiar SEO number into a misleading one. The fix is not a better rank tracker. Track whether AI answers cite you, then measure the visits, inquiries and revenue that follow. Retire position reporting for any query that AI answers touch.

The evidence that AI Mode sends fewer clicks

The strongest evidence so far comes from a preregistered randomized field experiment by researchers at the University of Pennsylvania and Northeastern University. Stephanie T. Wang, Jeffrey Gleason, Yakov Bart, Christo Wilson and Danaë Metaxa posted it as a draft paper titled AI in Search Reduces Publisher Referrals Without Improving User Experience in August 2026.

The design is what makes it hard to dismiss. 1,100 participants installed a browser extension that logged their real searches for seven days. Each person was randomly assigned to one of three conditions:

• Forced AI Mode: their Google searches were redirected into AI Mode.
• No AI: AI Overviews were hidden and AI Mode searches were sent to regular results.
• Current Search: Google behaved as normal.

In the preprint, assignment to AI Mode cut external click-through by about 19 percentage points per search compared with normal Google. Referrals fell to news sites, Reddit and Wikipedia alike. Hiding AI Overviews worked the other way: among participants where the hiding actually took effect, external click-through rose by 8.8 percentage points. AI Mode users also reported lower satisfaction and lower trust in information on Google, and they ran fewer search sessions per day. Removing AI features caused no detectable drop in satisfaction.

There are caveats. This is a draft that has not been peer reviewed. As Search Engine Journal's write-up notes, the extension's ability to hide AI Overviews broke partway through the study after Google changed its HTML. But AI Mode was enforced on 94.7% of searches in that group, so the click result for AI Mode is the cleanest finding in the paper. It is also the one that matters for planning, because AI Mode is where Google is taking search.

Why Search Console can't show you the loss

You might expect to see a drop like that in Search Console. Mostly, you won't, because of how Google counts these features.

Google's own help page, What are impressions, position, and clicks?, sets out the rules. An AI Overview takes a single position on the results page, and "all links in the AI Overview are assigned that same position." A click on an external link inside the Overview counts as a click, and standard impression rules apply. For AI Mode, Google says position "follows the same methodology as a Google Search results page."

Google's John Mueller has confirmed in public that Search Console treats an entire AI Overview as one block, as Search Engine Journal reported. In practice that block usually sits above the organic results. So a prominent citation card and a link hidden behind "Show more" can both report position 1.

That flattening breaks average position in two ways:

• It inflates you. Being one of a dozen sources in an AI Overview looks the same in the report as owning the top organic result, even though the click behavior is very different.
• It blends incompatible numbers. Search Engine Journal gives an example. A page that shows at position 1 when an AI Overview appears, and at position 19 when it doesn't, with equal frequency, reports an average position of 10. That page never actually ranks 10th. The number describes no real result page.

Google's newer Generative AI performance report helps a little. It shows how often links to your site appear in AI Overviews and AI Mode, broken down by page, country and device. But Google's documentation for it describes impressions only, and it aggregates them at the property level by default, so two of your URLs in one AI answer count once. It tells you whether you were present. It does not tell you whether you got the click.

Position was always a proxy. Now it is a broken one

Average position never mattered for its own sake. Executives accepted it because it tracked traffic closely enough, and traffic tracked pipeline. On queries that AI answers touch, both of those links are now weak. A position of 1 can sit alongside falling clicks, and the experiment above suggests that is the normal outcome as AI Mode use grows.

If your monthly report still shows a CMO a chart of average position, you are giving them a number that can improve while revenue from search declines. That is worse than having no number, because it invites bad decisions.

A measurement stack for AI search

Replace position with a stack of four layers. Each answers a question leadership actually cares about.

1. Citation presence

Treat AI visibility as yes or no. For a fixed set of priority queries, record whether your brand or URL is cited in the AI Overview or AI Mode answer, and which competitors are cited next to you. Search Console's Generative AI report gives you the page-level impression count. For the query-level view, sample the live answers on a schedule. An agent with a headless browser can run that weekly and log the cited domains to a sheet. Report share of citations across the query set, not a rank.

2. Visits and engagement from AI-affected queries

Split your Search Console queries into those that trigger AI features and those that don't, then report clicks and click-through rate for each group separately. Blending them is what produces averages that mislead. Pair the AI group with landing-page sessions in your analytics, so you can see whether the visitors who still click are better qualified than before.

3. Assisted conversions and inquiries

AI answers often do the research step, and the visit comes later through a branded search, a direct visit or a sales conversation. Look at multi-touch or assisted-conversion views for organic search, not just last-click. Add a "how did you hear about us" field that lists AI assistants as an option. Inquiries and revenue are the outcome. Everything else is diagnosis.

4. Branded demand

When an AI answer cites you but doesn't send the click, the effect often shows up as more people searching your name later. Track branded impressions and clicks in Search Console over time as a lagging indicator of AI visibility. If citation share rises and branded demand stays flat, the citations aren't persuading anyone.

What to stop reporting

Remove average position from executive dashboards for any query set that AI features touch. Stop reporting "number of keywords in the top 3" for those queries too, because flattening inflates the count. Keep position as an internal diagnostic for non-AI queries, where it still describes something real.

What to do about it

1. Tag your query universe this month. Split Search Console queries into AI-affected and non-AI groups, and report clicks and CTR for each group separately from now on.
2. Set up a weekly citation check. Pick 50 to 200 priority queries, log which domains AI Overviews and AI Mode cite, and report your citation share against your top three competitors.
3. Replace the position slide. In leadership reporting, use citation share, AI-group clicks, assisted conversions and branded demand. Keep position only for non-AI queries.
4. Rebuild attribution for delayed demand. Turn on assisted-conversion reporting, and add AI assistants as an option in your self-reported attribution field on every lead form.
5. Audit pages that are cited but get no clicks. Where the Generative AI report shows impressions but analytics shows little traffic, give people a reason to click that a summary can't replace, such as original data, tools or pricing detail.