What changed: the shift in how people find and choose a business

The feed stopped being a list of the people you follow.

This is the change everything else rests on, and it is worth knowing exactly how much of it the platforms have actually admitted to, because it is less than you would think.

"Right now, about 30% of the posts on Facebook feed are delivered by our AI recommendation system. That's up 2x over the last couple of years." And: "For the first time ever, more than 50% of the content people see on Instagram is now AI recommended."

Those are the only hard numbers Meta has ever published on this, and they are 28 months old. Meta has not restated either figure since. What it says now is qualitative:

"A very large percent of the content on Instagram and Facebook is not from your friends. It may not even be from people that you're following directly."

What the other platforms say

Instagram declines to publish the split. Its standing public explanation describes "a mix of content from the accounts you've chosen to follow, recommended content from accounts we think you'll enjoy and ads," and quantifies none of it. On Reels it will say only that "the majority of what you see is from accounts you don't follow." ⁹

TikTok's canonical explanation is six years old. It still states that "neither follower count nor whether the account has had previous high-performing videos are direct factors in the recommendation system." That is TikTok's stated design, published June 2020, and it should be read as that rather than as a description of the system today. ¹⁰

1. Meta Platforms, Q1 2024 earnings call transcript, 24 April 2024. Meta did not define whether "content seen" includes Stories, Reels and Explore. 1st
2. Meta Platforms, Q3 2024 earnings call transcript, 30 October 2024. 1st
3. Instagram, Instagram Ranking Explained , 31 May 2023. 38 months old. 1st
4. TikTok Newsroom, How TikTok recommends videos #ForYou , 18 June 2020. 74 months old. 1st
5. Meta Platforms, Q2 2026 earnings call transcript, 29 July 2026. 1st
6. No first-party platform use of "interest graph" or "interest media" was found, 8 August 2026. n/a

Two thirds of searches now end without anybody going anywhere.

This is the most measured change in the document, from the same clickstream panel across three readings, and it is worth showing as a series rather than as a single number.

Same analyst, two different data providers, three studies over two years. The 2024 readings come from the Datos panel and the 2026 readings from Similarweb's, which is worth knowing before treating the series as one continuous line. Neither publishes its panel size. ^{17 , 18}

And the traffic itself

From the same 2026 study: the share of traffic Google sends to the open web fell 8 percentage points between June 2025 and May 2026, "a roughly 22 percent drop in the traffic share Google sends." ¹⁸

What Google says about all this

"Total organic click volume from Google Search to websites has been relatively stable year-over-year." No number, no base, no comparison figure appears anywhere in that post. We are leaving the disagreement standing, because it cannot be settled from outside: Google is the only party with the data and it publishes none. ¹⁹

1. SparkToro with Datos, 2024 Zero-Click Search Study , 2 July 2024. Panel covers September 2022 to May 2024. 2nd
2. SparkToro with Similarweb, 9 and 16 June 2026. US and per-country cuts of one study, January to April 2026. Panel sizes not published. 2nd
3. Google, AI in Search is driving more queries and higher quality clicks , 6 August 2025. 1st

What people do before they decide, and how thin the evidence really is.

One well-documented model, one well-known survey, and a gap where the research about your actual business should be.

Google's "messy middle," with its real method attached

Google's own 2020 research programme, run by its UK market insights team. What it did: a literature review, several hundred hours observing 310 recorded shopping journeys across 31 categories, and a conjoint experiment across 31,000 in-market shoppers producing 310,000 simulated purchase scenarios. ²⁰

What it found: people loop between exploring and evaluating rather than moving down a funnel, and six cognitive biases can shift preference inside that loop.

What it did not test: actual purchases, actual revenue, local businesses, restaurants or independent operators. The experiment set a fictional challenger brand against a person's favourite brand and measured stated preference in a simulation. It is 72 months old. Applying it to a single-site restaurant is an extrapolation, and we would rather say that than let it pass.

Reviews, and the self-report problem

The most-quoted figures on social proof come from one annual consumer survey: 1,002 US adults, published 11 February 2026. It reports 97 percent read reviews for local businesses, 92 percent care about star ratings, and 68 percent require a minimum four-star rating, up from 55 percent a year earlier. ²¹

Two caveats belong next to every one of those numbers. It is entirely self-report, and asking somebody whether they read reviews reliably overstates it, because it is the socially expected answer. And the publisher sells reputation software. Jumps as large as 55 to 68 percent in a 1,002-person panel are as likely to reflect a changed question as changed behaviour.

Directionally the finding is old, consistent and safe to assert: reviews matter to a local decision. The specific percentages are not safe to assert as fact.

1. Google, Think with Google, Decoding Decisions: Making Sense of the Messy Middle , by Protheroe, Rennie, Charron and Breatnach, Google Market Insights UK, 2020. 1st
2. BrightLocal, Local Consumer Review Survey 2026 , 11 February 2026. 1,002 US adults, SurveyMonkey panel. US only. 2nd