AI Is Outrunning Our Measurement: The H1 2026 Trust & Attribution Gap, Decoded

Hi everyone, this is Neo.

Today I want to talk about something uncomfortable but genuinely important: AI’s impact has outrun our ability to measure it.

That’s not my line — it’s Kevin Indig’s. Who is Kevin Indig? Former Head of Growth at Shopify, former Director of Search Product at Gartner, now an independent advisor working with companies like Meta, Airbnb, and Reddit. Every six months he publishes an “AI Halftime/Year-End Report,” and the H1 2026 edition dropped a couple of weeks ago. Search Engine Journal covered it in depth almost immediately.

The report opens with one line that sets the thesis for the entire first half of the year:

“AI’s impact kept growing, but the ability to measure it kept falling behind.”

That gap between impact and measurement is the real story of H1 2026 — more than any single product launch or earnings call. In this post I’ve pulled out the most valuable data points from the report, layered on SEJ’s analysis and the UK CMA’s world-first regulatory ruling, and turned it all into a summary you can actually act on for the rest of the year.

Let’s start with the punchline: if your AI search monitoring still relies on a single rank-tracking tool, you’re probably seeing only about 9% of the truth.

1. The Scariest Number: 91% of AI Citations Appear on Just One Platform

Kevin’s research turned up a stat that I’d suggest every person doing AI monitoring print and tape to their wall:

91% of AI citations appear in only one of ChatGPT, Perplexity, or AI Overviews — never in more than one.

In other words: the fact that ChatGPT cites your page says nothing about whether Perplexity or Google’s AI Overviews will cite you too. There’s almost zero overlap between platforms.

What does that mean? Tracking just one platform is close to tracking none. In Kevin’s own words: “Tracking only one engine is nearly the same as tracking none.”

The reasons aren’t hard to find: every AI platform runs different models, different reasoning, different personalization, different update cadence — and the output even has stochastic variability. When measurement is this fragmented, the old “what rank am I” mindset simply stops working.

So how should you measure instead? Kevin’s answer: like a pollster, not a rank checker

This is the key methodological shift Kevin proposes:

Prompt tracking should behave more like polling and focus-group research than like the rank tracking SEOs have leaned on for two decades.

Traditional rank tracking is deterministic — you check once, you know the position, you check again tomorrow. AI search is probabilistic and nonlinear — ask the same question today and tomorrow and you may get different answers; change the user or their login state and the answer changes again.

So the right approach is: build a prompt panel spanning ChatGPT, Claude, AI Overviews, AI Mode, and at least one open-weight model. Run a fixed set of questions repeatedly, then read the results the way a pollster reads a sample — looking at distributions, trends, and proportions, not single snapshots.

Kevin’s exact framing: “Treat the results the way a pollster treats a sample rather than the way an SEO treats a SERP.”

2. Brand Mentions Matter More Than Citations

Here’s another counterintuitive but crucial finding from the report:

Brand mentions in AI answers correlate more closely with real business outcomes than citations do.

Most independent site owners are still measuring “how many times did AI cite me.” But Kevin argues citations are just the material that shapes an answer; what actually drives buyer behavior is whether your brand gets named, mentioned favorably, and recommended.

Here’s a practical framework you can copy straight from the report:

  • Citation: You’re listed as an information source — important for some businesses, but not the critical metric for most merchants
  • Mention: Your brand name appears in the AI answer — this is what most sellers should be tracking
  • Recommendation: You’re named ahead of competitors — the ultimate goal

The thing to monitor is: across your prompt panel, how often do you show up? In what context? Is the sentiment positive? Are you ranked ahead of your competitors?

SEJ’s analysis adds an even harsher detail: Google’s own Search Console data is reportedly about 75% incomplete for this new landscape. Even the “official data” only shows you a quarter of the picture — anyone still using GSC as their only AI measurement tool is working from a badly mangled puzzle.

One consumer-behavior stat in the report should be a shot of adrenaline for anyone doing brand work:

Roughly three out of four consumers pick the first result in an AI shortlist — unless a brand they already trust appears anywhere on that list, in which case they pick the trusted name instead.

In plain terms: in an AI shortlist, trust beats the default advantage of being first.

Another data point: 88% of the time, users accepted AI Mode product recommendations as “the best there is.” In AI Overviews, by contrast, users click, evaluate, and compare — classic search behavior. In chatbot scenarios, users barely compare. They just trust.

For independent site owners this means two things:

  1. Brand building has a newly measurable ROI — you could never quantify “brand equity” before; now you can: when AI lists five vendors, will the user pick you because they recognize you?
  2. Content has to be “mention-worthy” — why would AI name you? Kevin’s answer: lead with unique information, write in a direct, easy-to-understand way, cut the fluff, and keep your site technically fast and accessible. These fundamentals determine whether AI is willing to put you in the answer.

Kevin sums it up in one of the sharpest lines in the report:

“AEO/GEO is a brand channel disguised as a performance channel. AI recommendations shape demand, not citations. The unit of optimization is whether AI names, trusts, and recommends your brand.”

4. Meta Burned 73.7 Trillion Tokens in a Month: A Microcosm of the AI Bubble

My favorite anecdote in the report is also the most cautionary: Meta’s internal “Claudeonomics” leaderboard.

Here’s how it went down:

  • A Meta engineer built an internal leaderboard ranking 85,000+ employees by the number of AI tokens they consumed
  • Employees raced up the board by leaving AI agents running on idle or useless tasks — just to earn titles like “Token Legend”
  • Result: Meta engineers burned 73.7 trillion tokens in 30 days — and nobody could name the ROI
  • In April, CFOs slammed the brakes after discovering the annual token budget had been blown through in four months
  • The #1 user on the board averaged 281 billion tokens — over $1.4 million at standard API rates
  • The leaderboards were shut down in April, and the company mantra shifted from “TokenMAXXXING” to “valueMAXXXING”

The irony is thick: Jensen Huang started the year saying a $500K engineer who doesn’t consume at least $250K of tokens would “deeply alarm” him — and companies responded by proving that mindlessly burning tokens doesn’t equal creating value.

But there’s another side to the coin

Kevin honestly notes the token tsunami wasn’t all waste:

  • The wave of usage around Claude’s Opus 4.5/4.6 releases pushed far more people into daily AI use than any marketing campaign could have
  • That surge flowed down the stack to infrastructure companies (like Supabase, which Claude kept recommending) — second-order effects that never show up on an earnings call but shape an entire ecosystem

Neo’s take:

The lesson for independent site owners: “how much AI you used” and “how much value it created” are two different things. Don’t copy Meta and chase volume. When your team uses AI, ask “how many hours did it save me, how much money did it make,” not “how many tokens did I burn today.” Track value, not consumption.

5. Software Stocks Fell 30%: The Market Priced Narrative, Not Fundamentals

Another observation worth chewing on: software stocks fell close to 30% over the first half.

But Kevin’s data shows the decline tracked almost entirely with how the market perceived a company’s AI disruption risk, not with how the company actually performed:

  • The bottom quartile of software stocks dragged the whole sector down
  • The median and top quartile actually outperformed the broader ETF
  • In other words: this was a selloff driven by narrative, not fundamentals

The “AI layoffs” story is equally distorted. Challenger, Gray & Christmas reported more than 87,000 job cuts cited AI as the reason through May — roughly a fifth of all 2026 layoffs.

But Kevin’s tracking (which goes back to his H1 2025 predictions) shows: AI layoffs are mostly a PR narrative. The real causes are pandemic-era overhiring, capital expenditure discipline, and economic turbulence. Even better: some companies that blamed AI for layoffs quietly started rehiring.

Neo’s take:

For cross-border sellers, this is an information arbitrage: while overseas media screams “AI replaces everything,” the real business world is far less linear. Don’t let the narrative drive your decisions — whether it’s panic pivoting (abandoning all traditional SEO) or opportunistic FOMO (blindly all-in on some new trend), come back to the data. Your real competitor isn’t AI. It’s panic.

6. The Model Landscape Shook Up: ChatGPT’s Share Fell From 78% to 56%

The report also contains a landscape shift that everyone doing AI optimization must know:

Platform July 2025 July 2026
ChatGPT 78% 56%
Gemini 15% 30%
Claude 2% 10%

In one year, ChatGPT lost 22 points of share, Gemini doubled, and Claude grew 5x. Kevin’s read: Google is now the most likely winner of the AI consumer market — ChatGPT has lost its significance in AI Search to Google.

A few other notable details:

  • OpenAI is pivoting to enterprise (enterprise is ~40% of revenue, heading toward 50% ahead of a potential IPO) and killed Sora, video in ChatGPT, and Instant Checkout
  • Per Ed Zitron, OpenAI spends $2.8 billion a month to make $1.1 billion — an alarming burn rate
  • Open-weight models (Kimi K3, GLM 5.2, DeepSeek V4) are pressuring the US labs
  • The big labs massively subsidize tokens: SemiAnalysis found a $200 subscription delivers $8,000–$14,000 worth of tokens

Neo’s take:

Model choice has become a genuine business risk. Last year everyone optimized only for ChatGPT; this year you must cover Gemini and Claude too — because users are migrating. This is also why the multi-platform prompt panel from section 1 matters so much: you don’t know which assistant your users will use next year, so you need to be visible in all of them today.

7. The Regulatory Storm: Publishers Take the Fight to the Courts

Everything above is business-level. This section is institutional — and every item directly shapes the future of the content ecosystem:

1. Munich Court: Google Liable for False Statements Generated by AI Overviews

In June 2026, a court in Munich ruled Google liable for false statements generated by AI Overviews. A global first — the platform carries legal responsibility for AI-generated errors.

2. 400 Newspapers Sue OpenAI and Microsoft

In June 2026, 400 newspapers jointly sued OpenAI and Microsoft over unauthorized use of their content.

This is the one I think has the deepest implications for independent site owners. On June 3, 2026, the UK’s Competition and Markets Authority imposed a world-first conduct requirement on Google Search, including:

  • Publishers can prevent their content from powering AI search features (like AI Overviews)
  • Google must properly attribute publisher content in AI-generated results, with clear links
  • Publishers can opt out of their content being used to fine-tune AI models
  • Google has nine months to implement, then must submit compliance reports every six months

SEJ’s analysis adds important context: USA Today’s CEO went as far as saying “We’re getting close to the point where we’ll block Google and abandon the traditional search traffic we get today.”

Neo’s take:

Three implications for independent site owners:

  1. Attribution = a free traffic channel: when Google is forced to attribute clearly, being cited becomes more valuable — provided your content is actually worth citing
  2. Content licensing is becoming a real business: publishers are shifting from “content-for-traffic” to “content-for-training-license-fees.” The value of your site’s content to AI crawlers may one day be monetizable
  3. Get compliance-savvy early: these rulings are European, but Google operates globally (with some exceptions). Understanding AI attribution and opt-out mechanics will be useful when negotiating or defending your rights later

8. The H2 2026 Core Thesis: Separating Intelligence From Agency

At the end of the report, Kevin drops a prediction for the second half that I’d suggest you read three times:

“H2 2026 will separate intelligence from agency.”

What does that mean?

  • Intelligence is depreciating: with open-weight models and the token price war, model capability keeps getting cheaper and more commoditized — being “smart” is no longer scarce
  • Agency is appreciating: what’s getting locked down is the permission to act on someone’s behalf — spend their money, access their data, act as them — and it’s being tightened by platforms, governments, payment networks, and users themselves

The evidence is already here: the publisher lawsuits and the UK CMA ruling are early proof that this tightening has started.

What does this have to do with SEO strategy? Everything. Here’s Kevin’s practical to-do list for H2, passed through my own lens:

Action 1: Retire the single rank-tracking tool mindset

Build a prompt panel spanning ChatGPT, Claude, AI Overviews, AI Mode, plus at least one open-weight model. Read results like a pollster, not a SERP checker.

Action 2: Shift reporting from citation counts to mentions + sentiment + recommendation rank

A dashboard that only counts citations measures the smaller half of what actually drives buyer behavior and will keep understating your real AI footprint to clients or leadership. Switch to: mention frequency, sentiment, and recommendation rank.

Action 3: Audit your agentic access layer now

This is the most forward-looking item in the report, and the one independent site owners can act on immediately:

  • Is your product data, pricing, and checkout flow structured so an authorized AI agent can complete a transaction on a customer’s behalf?
  • When an AI agent is choosing among several competitors, are you the trusted, nameable option it’s permitted to act for?

Kevin’s call: the winners in H2 won’t be the brands that get mentioned most often — they’ll be the brands agents are permitted to act on behalf of.

9. Neo’s Summary: An Action List for Independent Site Owners

Let me compress the whole report into five executable recommendations:

  1. Rebuild your measurement system: stop looking at a single rank-tracking tool; build a prompt panel across ChatGPT/Gemini/Claude/AIO and analyze it with a pollster’s mindset (you can start this week)
  2. Upgrade your KPIs from citations to mentions + recommendations: monitor how often your brand is named in AI answers, in what context, with what sentiment, and whether it’s ranked ahead of competitors
  3. Trust over traffic: 3/4 of users pick the first result and 88% accept AI Mode recommendations — a trusted brand is the only force that breaks the “first is default” advantage. Keep investing in brand content
  4. Cover multiple models: ChatGPT lost 22 points of share in a year while Gemini and Claude rose — you need to be visible in every assistant
  5. Prepare for the agent era: structure your product data, pricing, and checkout so your brand becomes something agents are willing to act for

One honest closing thought:

Kevin ends his report agreeing that “the ability to measure AI’s impact fell behind the impact itself in H1 2026” — and he doesn’t think that gap closes on its own by year-end. I agree. Measurement tools will always chase the pace of technology.

But that’s exactly the opportunity: when most people are measuring the new world with an old ruler, switching to a new ruler first means you see what others miss — and you act first.

AI may be outrunning measurement. But your actions can outrun AI.