Trust is the AI era’s currency. Five viral stats that fail a source check

Why trust became the most expensive asset of the AI era

Producing content became almost free in 2026. According to Graphite’s May 2026 study, roughly half of all newly published web articles are primarily AI-generated, and that share has held steady for five consecutive quarters. When volume no longer proves anything, the differentiator moves somewhere else: whether what you say can be verified.

Trust in the AI era does not mean “people trust us”. It means every claim on your site leads to a source the reader or the model can open — and that you correct yourself in public when the source proves you wrong.

Google says the same thing in its own words. Its Search Central guidance (updated 10 December 2025) states that of the four aspects of E-E-A-T, trust is the most important, and that the other three exist to support it. The reader-side numbers are uncomfortable too: in a Pew Research Center survey (5,153 US adults, fielded 18–24 August 2025, published 1 October 2025), only 6% said they trust the information in search AI summaries “a lot”, and 46% had little or no trust in it. The clarification that usually gets lost: those percentages are of people who have actually encountered AI summaries, not of the whole sample.

And here is the awkward part. A great deal of what gets repeated about AI search on LinkedIn and in industry blogs does not survive the same check those articles demand of their readers. We took the five most common numbers and traced each one to its origin. The results are below — including the one place where we were wrong ourselves.

Stat 1: “94% of B2B buyers start their research with AI”

The number is real. The word “start” is not.

The primary source is 6sense’s 2025 Buyer Experience Report (press release 12 November 2025, more than 4,000 B2B buyers across North America, EMEA and APAC). It contains two separate 94% figures: 94% of buying groups had ranked their preferred vendors before first contact with any vendor, and 94% of buyers used large language models to summarise reviews or analyse data.

Neither says buyers start with AI. One section of that same press release is headed “LLMs Are Tools — Not Decision-Makers”, and 6sense’s own analysis of 2 December 2025 states that buyers use language models in the middle of the journey, not at the beginning or the end, and that the model works as a research assistant rather than a research initiator. The reason is in the same report: 85% of buyers had prior direct experience with the vendors they evaluated. They do not need ChatGPT to tell them who the players are.

If you want to support the claim that AI shapes the shortlist, better-documented sources exist. G2’s Answer Economy report (survey of 1,076 B2B decision-makers, fielded March 2026, published 15 April 2026) found that 51% of software buyers now start research with an AI chatbot more often than with Google, and 69% said a chatbot led them to a different vendor than the one they had planned on. Two honest caveats: G2 is a review platform whose commercial interest matches its conclusion, and this is self-reported survey data rather than observed behaviour.

TrustRadius’s 2026 report (1,862 buyers, surveyed January 2026) adds the number that sums up this whole story: 94% of those who used AI during the buying journey fact-check the AI’s answers at least some of the time. People do not take the machine at its word. They verify — and to verify, they come to your site.

Stat 2: “80% of shortlisted brands are decided in the first AI answer”

We could not find a primary source for this one. Not in the 6sense report, not at G2, not at Gartner.

The closest real figure comes from that same 6sense press release: the vendor a buyer preferred before ever talking to sales goes on to win 80% of deals, and 77% of buying groups did buy from their initial favourite. That is a claim about deal outcomes. Somewhere along the way it became a claim about positions inside an AI answer — two entirely different things wearing the same number.

This is how numbers travel. Nobody lies deliberately: someone paraphrases one sentence, the next person cites the paraphrase, the third puts it on a slide. Three steps later the number has authority and no source. The same pattern repeats throughout this article.

Stat 3: llms.txt — and where we got it wrong ourselves

Through 2025, SEO circles spread the advice to add an llms.txt file to your site root, summarising your content for language models. The logic sounded convincing. The one thing nobody asked was whether the models read the file at all.

Google’s own documentation now answers directly. Its guide “Optimizing your website for generative AI features on Google Search” (Search Central, updated 10 July 2026) contains a section titled “what you don’t need to do”, and the first item in it is exactly this file: Google Search does not use llms.txt files or any other “special” markup, and if you keep one anyway it will neither harm nor help your visibility, because Google simply ignores it. Google’s John Mueller had said the same on 17 June 2025: no AI system currently uses llms.txt, and it is obvious from server logs — the chatbots fetch your pages, but nobody fetches that file.

We also checked by hand on 3 August 2026. docs.anthropic.com/llms.txt responds (via a redirect to platform.claude.com/llms.txt), but anthropic.com/llms.txt, developers.google.com/llms.txt and platform.openai.com/llms.txt all return 404. Publishing the file and consuming the file are two different things.

Correction. Our article of 16 June 2026, “Generative Engine Optimization”, listed llms.txt both in its checklist and in its FAQ with the claim that it “helps models understand your content correctly”. That was the general industry understanding at the time, but in the light of Google’s 10 July 2026 documentation it is not true. We corrected that article on 3 August 2026 in all three languages. The file at iweb.ee/llms.txt stays — Google states explicitly that it does no harm — but we no longer sell or recommend it as an AI visibility measure.

That same section — Google titles it “Mythbusting generative AI search: what you don’t need to do” — dismisses four other widespread recommendations, and they are worth knowing before somebody invoices you for them. Structured data is not required for generative AI search and there is no special schema.org markup you need to add. “Chunking” your content for models is not required. Rewriting your content specifically for AI systems is not required. And chasing inauthentic mentions across the web is not as helpful as it seems. There is one more sentence on that page worth remembering whenever a tool offers you an AI visibility score: no third-party tool has access to Google’s internal ranking or AI systems.

The conclusion that follows is uncomfortable for our whole profession. Structured data and a clean technical foundation are still worth having — they help in classic search and make a page machine-readable — but the reason they are usually sold with is often wrong. We keep schema markup on every page because it works in ordinary Google search and rich results. Not because ChatGPT demands it.

Stat 4: “AI Overviews killed organic clicks”

The problem here is not bad data. It is that serious studies give different answers and almost nobody cites them together.

• Pew Research Center, 22 July 2025. 900 US adults whose browsing was tracked during March 2025: 68,879 Google searches in total. When an AI summary was present, users clicked a result 8% of the time; without one, 15%. A link inside the summary was clicked 1% of the time. The strongest source on this list — independent and non-commercial. But it is correlational, not causal: AI summaries appear disproportionately on informational queries that already converted poorly.
• Agarwal (Indian School of Business) & Sen (Carnegie Mellon), SSRN 2026. The only randomised experiment in this field: users were randomly assigned to Google with AI Overviews switched on or off. The result, conditional on an AI Overview actually appearing, was 39.8% fewer outbound organic clicks and 34.5% more zero-click searches, with no measurable improvement in how users rated search quality. It is a working paper, not yet peer-reviewed — and that has to be said every time the number is used.
• Semrush, 15 December 2025. Compared, before and after, the very keywords that did not trigger an AI Overview in May but did by October (the broader 200,000+ keyword dataset is the other half of the same study). The result cuts against the others: the zero-click rate fell, from 33.75% to 31.53%. In Semrush’s own words, that suggests AI Overviews do not automatically increase zero-click behaviour. Methodologically it is one of the cleanest comparisons available, and it comes from a vendor whose commercial interest points the other way.
• Seer Interactive, 24 April 2026. 53 brands, 5.47 million queries: the decline did not merely stop, it partly reversed — organic click-through on AI Overview queries rose from a December 2025 floor of 1.3% to 2.4% in February 2026. Seer itself warns against forecasting a recovery from two months of data.

Then there is the denominator problem. “AI Overviews appear on X% of searches” means something different every time: Pew measured 18% of real searches (22 July 2025), Semrush 15.69% of keywords (15 December 2025), Ahrefs 20.5% of SERPs (10 November 2025), BrightEdge around 48% of tracked queries (12 February 2026). Same phenomenon, close to a threefold spread — because they are counting different things. Any percentage without its denominator is useless.

An honest closing note: Google itself claimed on 6 August 2025 that traffic is stable and clicks are “higher quality”, dismissing third-party studies as having “flawed methodologies” — without naming a single one and without publishing its own data, chart or methodology. That missing evidence is a checkable fact in its own right.

Stat 5: “Half the internet is already written by AI”

This number is actually measured rather well — and almost always cited wrongly.

Graphite, May 2026. A sample of 55,400 randomly selected English-language Common Crawl URLs, classified by three separate detectors (Pangram, Copyleaks, GPTZero) and averaged. The result: primarily AI-generated articles accounted for 49.6% of new articles in Q1 2025, 50.9% in Q4 2025 and 49.9% in Q1 2026. The share has sat at roughly half for five consecutive quarters — not exploded, as it is usually described. Graphite also states its own limitation: it did not evaluate how accurate the AI detectors themselves are.

Far more common is a reference to the older figure — “51.7% by May 2025” and “AI overtook humans in November 2024”. Graphite has formally marked that study as superseded: it used a single detector, and the newer method puts the share 3.3 points lower. An article citing that number today is citing a study its own author has crossed out.

And then there is Ahrefs’ 74.2% (19 May 2025, 900,000 pages), which often gets placed alongside as a contradiction. It answers a different question: what share of new pages contain at least some AI content. Of those, 2.5% were pure AI, 25.8% pure human and 71.7% a mix. “Contains AI” and “is mostly AI” are not the same claim.

So does AI content lose in search? Here two serious studies genuinely conflict, and the honest answer is to say so. Graphite (October 2025, 31,493 keywords) found 86% of articles ranking in Google are human-written and only 7% of first-place results are AI-generated. Ahrefs (14 July 2025, one million SERPs) found the opposite: AI Overviews cite AI-touched content more than the web average. Anyone telling you the answer is settled has not read both.

What actually holds up: Google’s own words and the correlations

Strip out the unverified numbers and what remains is a surprisingly simple picture — and it supports exactly the conclusion the inflated versions were selling.

• Google says trust matters most, in its own words. Search Central guidance (updated 10 December 2025): of the aspects of E-E-A-T, trust is the most important and the others contribute to it. On the same page: E-E-A-T is not itself a ranking factor, but Google uses a mix of factors that identify content with good E-E-A-T. And: Google strongly encourages adding accurate authorship information.
• Signals outside your own site correlate with AI visibility more strongly than anything on it. Ahrefs’ analysis of 75,000 brands (12 December 2025) found the strongest association with YouTube mentions (Spearman ≈0.74), then branded web mentions (0.66–0.71), then branded search volume (0.35–0.47) — with domain authority (0.27–0.33) and page count (≈0.19) at the bottom. Ahrefs states plainly that correlation is not causation. And Google separately warns that chasing inauthentic mentions does not help. Read together: mentions are earned, not bought.
• Reviews remain the cheapest trust investment there is. BrightLocal’s 2026 survey (11 February 2026, 1,002 US adults — so a direction, not an Estonian number): 97% read reviews of local businesses, 47% will not use a business with fewer than 20 reviews, 74% only care about reviews from the last three months, and in BrightLocal’s own wording the most commonly used sources for local recommendations are Google, Facebook and AI tools like ChatGPT. The practical side is spelled out in our guide to local SEO for small business.
• Demand for transparency exceeds supply. Fractl’s 2026 study (1,008 US consumers and 150 marketers): among consumers, 84% want AI-written text to be labelled, while on the marketer side only 20% of organisations always disclose AI use and 33% never do. From the same study: the share of consumers who say heavy AI use would decrease their trust in a brand doubled in a year, from 20% to about 40%.

The honest counterweight, usually left out of articles like this one: transparency has a cost. In experiments by the Nuremberg Institute for Market Decisions (representative surveys of 1,000 respondents each in the US, UK and Germany, with the first experiment embedded in them), labelling an advert “AI-generated” made people rate it as less natural and less useful and reduced their willingness to buy — with identical creative. Two clarifications: that particular result comes from a second experiment with German participants only, and the institute itself describes the effect as small, though meaningful. Transparency is not a free win. It is an investment that pays back on the returning customer, not on the first click. Since 2 August 2026 it is also partly a legal obligation in the EU: the AI Act’s transparency rules took effect the day before this article was published.

Estonia: 46.6% versus 0.45%

The most telling Estonian number here is the gap between two verifiable sources.

Eurostat (16 December 2025): in 2025, 46.6% of residents of Estonia aged 16–74 used generative AI tools — only Denmark is ahead (48.4%) and Malta is just behind (46.5%), against an EU average of 32.7%. For work, the figure is 25.1% (Eurostat dataset isoc_ai_iaiu; the EU average for work in the December release was 15.1%). The neighbours are clearly lower: Lithuania 36.9%, Latvia 33.4%.

StatCounter (July 2026): of the referral traffic StatCounter measures in Estonia, 83.51% comes from search, 16.05% from social and 0.45% from AI chatbots. Within that 0.45%, ChatGPT accounts for 78.21%, Gemini 12.45%, Perplexity 3.36%, Copilot 2.74% and Claude 3.24%.

The two numbers are not in conflict — they measure different things. 46.6% is human usage; 0.45% is clicks that reach your site. Every conversation in which a model answers a customer about your industry and the customer clicks nowhere falls outside that 0.45%. So it is a floor, not a ceiling — which is precisely why the only metric that matters right now is whether the AI answer mentions your brand at all.

On the company side, the Estonian picture looks different. According to Statistics Estonia’s ICT survey (15 September 2025), 22% of Estonian enterprises with 10 or more employees used at least one AI technology in 2025 — up from 14% in 2024 (derivable from the same release) and 5% in 2023 (Statistics Estonia table IT149). The most common is the text generator (15% of enterprises). Among large enterprises the figure is 53%; among companies with fewer than 50 employees, about a fifth.

And here is the gap nobody mentions. All of that enterprise statistics starts at ten employees. Statistics Estonia recorded 159,827 enterprises in its 2025 statistical profile, of which 152,205 have fewer than ten employees — 95.2% (our calculation from their published figures). In other words, most Estonian business owners reading this appear in none of the percentages above. When somebody’s slide says “22% of Estonian companies use AI”, it is worth asking who that is about.

One more figure from Statistics Estonia at the same time (16 September 2025) sums up the point of this article: only 30% of respondents have verified the accuracy of information found on news portals or social media. Meanwhile Google holds 85.08% of the Estonian search market (StatCounter, July 2026) — classic search is not going anywhere, and the fundamentals of SEO still work.

The three-minute check: how to verify a number yourself

Writing this article took considerably longer than three minutes, but checking one number usually does not. Seven steps, in the order we used them above:

1. Find the primary source, not the site quoting it. Look for the publisher’s own page: the research firm’s press release, the statistics office table, the academic paper. If all you reach is blogs citing each other, there is no source.
2. Read the sentence the number sits in. “94% used language models during the buying process” and “94% start with a language model” are two different claims — the gap opened up by swapping one verb.
3. Check the denominator. A percentage of what? Searches, keywords, SERPs, users? For the same phenomenon we saw close to a threefold spread caused by the denominator alone.
4. Check the date, and whether the author has retracted it. Graphite marked its own study superseded. We corrected our own article. This happens more often than people assume.
5. Ask who benefits from the number. Most AI search data comes from tool vendors whose product is exactly what the finding implies you should buy. That does not make the data wrong — it means you read it with a different eye.
6. Look for someone who disagrees. Semrush disagrees with Ahrefs, Graphite disagrees with Ahrefs, Google disagrees with everyone. Citing one study where four exist is not citation, it is selection.
7. If three minutes did not turn up a source, don’t repeat the number. That is the whole method. Everything else is applying it.

The same logic applies to the proposals you receive for a website or for SEO. If someone promises a guaranteed position in ChatGPT’s answers, ask what metric they measure it with and where their data comes from — Google states plainly that no external tool has access to its AI systems. If no answer arrives, three minutes have saved you more than reading this article cost.

The practical side — what to actually do on your site so AI answers can cite you — is covered in a separate article on generative engine optimization (corrected on the llms.txt point on 3 August 2026). If you are weighing up whether to build the site with AI yourself, see our comparison of AI website builders and a web agency. And if you would rather someone built the technical foundation and the content systematically, there are iweb.ee SEO services — with solid website development as the precondition for all of it, not the consequence.

FAQ — frequently asked questions