If your models include search, paid (PPC) or organic, you may well have noticed that performance appears to be changing. Several organisations I work with – in UK services, UK retail and European gaming – have seen paid search clicks fall by a quarter to a third over the last year. Organic search traffic has fallen even further in some instances. Yet in most cases, online sales have held flat (or nearly so).

That should worry anyone building a Marketing Mix Model (MMM) on KPIs like online transactions, new customer registrations or simple online sales. And more change is coming. Google is bringing ads into its AI search answers in the UK this quarter, and I understand ads in Gemini will arrive towards the end of Q1 2027.

This piece is a heads-up: what has changed, what is coming and what I’d do about it when building and reporting on a model.

More searches now end on Google itself. SparkToro’s 2026 study found that 68% of US Google searches in January–April 2026 ended without a click to any website, up from about 45% a decade ago. People ask a question, read Google’s AI-generated answer and stop there.

Other studies point the same way. Pew Research tracked the browsing of 900 US adults: when an AI summary appeared, people clicked a traditional result in 8% of visits, against 15% when it didn’t. Seer Interactive found paid click-through rates on informational searches with an AI Overview fell 68% between June 2024 and September 2025.

The pattern in my client work matches. The decline started in late 2025 or early 2026, depending on the category, and runs as a steady downward line, slowing for some. Seer’s 2026 update suggests click-through rates on some searches may be levelling out, but where I see a slow-down, the decline hasn’t stopped.

Most of the lost clicks are being kept by Google, not taken by ChatGPT and other AI tools. In SparkToro’s analysis of US desktop search across 41 major websites in Q4 2025, Google handled 73.7% of searches and AI tools 3.2%. People are changing how they use Google as much as where they search.

Search has fallen, online sales haven’t (illustrative data)
Search has fallen, online sales haven’t. From late 2025 search volume falls to about 82, paid search clicks to about 73 and organic clicks to about 60, indexed to 2024 = 100, while online sales hold near 100. Decline begins,late 20255060708090100110Jan 2024Jan 2025Jan 2026Organic clicksPaid search clicksSearch volumeOnline salesWeekly, seasonally adjusted, 4-week average, indexed to 2024 average = 100. Illustrative data.

What is coming

Google is now putting ads into these AI answers. Ads in AI Mode – Google’s conversational search tab – reached the UK in Q4 2026, alongside Brazil and Mexico. In the US, Google is also adding ads to the follow-up questions within an AI Mode conversation.

The Gemini app is next. Google spent late 2025 and most of 2026 first denying, then softening, its position on Gemini ads. On its Q3 2026 results call on 28 October, it said its Gemini agent is coming to consumers, starting with the Gemini app, and that it is confident it can build advertising into agents – its example was a sponsored offer shown alongside the agent’s own recommendations. Ads in Gemini are believed to arrive towards the end of Q1 2027.

We don’t yet know how these placements will behave. A sponsored option inside a conversational answer is not the same thing as a text ad above a list of links. It may be seen without being clicked. It may reach people earlier or later in their journey. Google has said ads in AI Mode earn less than ads in classic search. And who knows how ad blockers will be able to respond?

For marketing mix modellers, the key question is this: will what we have learned about paid search response in our models carry over? It seems unlikely.

Why this matters for your model

Search has always been an awkward variable to model, whether paid or organic. Most search ads are triggered by someone who is already looking for you (or at least the category), so ad volume rises and falls with demand – and demand is what the model is trying to explain. Put paid search clicks into a model unadjusted and they soak up a large share of the variation in sales. Much of that credit really belongs to the brand, to other media channels, to promotions and to underlying demand.

This is the endogeneity problem, and modellers have spent years working round it. The usual remedies are a control for underlying search demand, or a secondary model that explains paid search from the other drivers before it enters the main model. None of this is new: we were advising clients to do it before Google’s own 2018 paper on bias correction for paid search set out the logic.

What’s new is that the link between search and sales is breaking apart in new ways. When paid and organic search fall by a third and sales barely change, a model that has previously learned “more search clicks, more sales” will expect sales to fall. They don’t. Something else has to fill the gap: the base, a trend or, worse, another media channel that happened to grow over the same period. That is where bias creeps into the rest of the model. Get one variable wrong and the credit may go elsewhere by mistake.

The flat sales are also evidence in their own right. If a third of the clicks can disappear without denting sales, those clicks were doing less than many models gave them credit for – which is the endogeneity problem showing itself in the data.

Why paid search takes the credit
Why paid search takes the credit. Consumer demand drives searches, clicks and sales, and also reaches sales by other routes. Because clicks and sales move together, the model credits search. Now 68% of US searches end in an AI answer without a click, so clicks fall while sales do not, and the gap goes to the base, a trend or another channel. ConsumerdemandGooglesearchClick toyour siteOnlinesaleDemand reaches sales by other routes tooAI answer,no click68% of US searchesend without a clickThe model sees clicks and salesmove together – and credits searchNow clicks fall and sales don’t.The gap goes to base, trend – or a channel.

Search volume is a weaker control than it was

Google’s MMM guidance recommends Google Query Volume (GQV) – similar to Google Trends, but designed for MMM – as a control for paid search. The logic is sound: hold underlying demand constant and what remains is closer to the effect of the ads.

That logic assumes searches and site visits move together. They no longer do. Google reported its fastest year-on-year growth in search queries for five years in Q3 2026, while click-through keeps falling. More searches, fewer visits. Those extra searches are increasingly longer, conversational questions, so total queries can rise while searches for your own brand and category terms fall – as they have for the organisations I work with. A control that tracks searches is drifting away from the behaviour it was meant to stand in for. There is no perfect fix and each organisation is different, so test whether the relationship still holds in your own data rather than assuming it.

There is a second problem. As far as I know, Google doesn’t at present report ads in AI Overviews or AI Mode separately; they are folded into standard campaign reporting. A “paid search clicks” series in 2027 will mix placements that behave differently, with no visible break in the data. The variable will change meaning without telling you.

What I’d do

There isn’t a perfect answer, and each organisation is different. As always, start by visualising the data. Then try the simpler approaches below, and move to more advanced methods only if you need to.

  1. Visualise before you model. In any data-led investigation, this comes first. Plot weekly search volume, paid and organic clicks and online sales over the last three years, with brand and generic kept apart. Look at the ratios too: click-through rate (clicks per impression), cost per click and cost per thousand impressions. A sudden shift in a ratio usually means the nature of the channel has changed, rather than demand – a move into Performance Max is a typical cause – and it’s this kind of check that tells you when it happened. In the work I’ve seen, organic search volume is not holding up and clicks are falling. Some of the lost organic traffic does take sales with it, but nowhere near at the rate search has fallen.
  2. Not all impressions are equal. Advertisers using Performance Max (PMax), Google’s automated campaign type, will often find their accounts full of cheap impressions. I’ve seen cost per thousand impressions fall to a quarter of its former level for some heavy users of PMax, although that situation is fluid. PMax itself isn’t the problem: in UK furniture studies I’ve worked on, it delivered similar returns to non-PMax activity at around 10% lower cost, a result we validated both in the mix model and outside it. The key point is that not all impressions are equal. A PMax impression is fundamentally different from a search ad impression, and your work needs to recognise that.
  3. Choose the paid search measure with care. We’ve always favoured impressions for paid search, because they measure what people actually saw – their exposure to the ads. Some practitioners use clicks, but we’re sceptical: clicks are often simply part of the sales journey, so they rise and fall with demand. But not all impressions are equal (point 2), so spend is worth a look. It can be a useful measure, and the less aggregated the data series, the more sense it makes. A single paid search total is a basket of very different campaigns – brand, generic and competitor terms, and beneath those, different product families and campaigns for retention, customer service, awareness or consideration – each with its own results. This is the familiar basket problem in econometrics. Spend goes part of the way to solving it, and it reflects bidding changes as well as search volumes. Better still, split the basket where the data allows: in most cases, model brand and generic paid search separately at the very least.
Which paid search measure goes in the model? (illustrative data)
Which paid search measure goes in the model? For one account, indexed to 2024 = 100: clicks fall steadily from late 2025; impressions jump to more than three times their earlier level when Performance Max floods the account; spend is steadier, falling only slightly. ClicksPart of the sales journey,so they follow demand100200300400202420252026ImpressionsJump when PMax floodsthe account202420252026SpendSteadier; reflects biddingas well as volume202420252026One account, weekly, 4-week average, indexed to 2024 average = 100. Shaded: after the shift. Illustrative data.
  1. Test your search volume control. If you control for search volume, check whether it still tracks site visits and sales in your own data. In many cases that relationship is coming apart.
  2. Model the shift, simply at first. A before/after flag from the point the decline began – late 2025 or early 2026, depending on the category – is a good start. This is a step dummy: 0 before the change, 1 after. Add a trend term from the same point, counting up 0, 1, 2, 3… each week: the decline hasn’t bottomed out yet, so most organisations I work with need at least a linear trend, and it works well enough in most cases. Move the start date a few weeks either way and check the other media results barely change. If they swing about, the trend is taking credit from them, or the other way round. And be careful extrapolating it: a linear trend carried into a forecast or budget scenario keeps falling for ever, so decide explicitly where it levels off.
Last year’s model rolled forward vs refitted with a flag and trend (illustrative data)
Allow for the shift and the errors settle. Left: last year’s model, rolled forward, under-predicts sales from late 2025 and its errors grow steadily one way. Right: refitted with a before/after flag and a trend, the model tracks sales and the errors settle around zero. Last year’s model, rolled forward80100120ErrorJan '24Jan '25Jan '26Refitted with a before/after flag and trendJan '24Jan '25Jan '26Actual salesModelShaded: after the shift. Illustrative data.
  1. Put the same trend into any secondary model. If you model paid search from other drivers before it enters the main model, that secondary model must account for the same shift in behaviour. Otherwise it will blame the fall on whatever drivers it contains.
Put the trend into the secondary model too
Put the trend into the secondary model too. Stage 1 explains paid search from TV and other media, search demand, seasonality and a trend from late 2025. Stage 2 explains online sales from the part of paid search stage 1 does not explain, other media, price and promotions, and the base, flag and trend. Without the trend, stage 1 blames the fall on whatever drivers it holds. Stage 1 · paid search modelStage 2 · sales modelTV and other mediaSearch demandSeasonalityTrend from late 2025PaidsearchPaid search notexplained by stage 1TV and other mediaPrice and promotionsBase, flag and trendOnlinesalesLeave the trend out of stage 1 and it blames the fall on whatever drivers it holds.
  1. Go further if you need to. Time-varying parameters let the paid search effect drift over time. A state-space model, estimated with a Kalman filter, can do the same for the base, letting the intercept drift rather than forcing a straight-line trend. Both need care: the more freedom the base has, the more media credit it can absorb.

The usual hold-out tests for paid search still work, and are worth running if you can. But the change is in how consumers use Google, and you can’t switch off organic search – which is where the P&L is being hit hardest. Because the shift is happening in every territory your models cover, regional testing won’t isolate it. For the new AI Mode and Gemini placements, the practical step is to mark their go-live dates in your market now and look for a change in paid search response at those points. Do the same for organic search: new search engines and AI assistants taking share, and changes to how Google weights its results. Your SEO team or agency should be able to tell you when these happened, so you can mark them in the same way.

Talk to The MMM Doctor

If this has affected your models and you’re not sure where to start, come and talk to us. We can point you in the right direction, work with your modellers if they haven’t yet thought this problem through and make sure you’re ready to model through the changes coming next year. The aim is the same as always: reliable, robust models with as little bias as the data allows.

Has paid search started to behave differently in your models? Get in touch.

Sources