AI halftime report: H1 2026

Every six months, I return a crisp look astatine the latest developments successful AI and Search. Based connected really quickly things are developing, I would almost request to summation that cadence to monthly. (And possibly I will.) But first, here’s what happened successful H1 2026.

H1 2026 moved money, traffic, jobs, and marketplace headdress earlier anyone could beryllium really overmuch worth AI created. Search behaviour changed, token budgets exploded, package stocks sold off, and companies blamed layoffs connected AI.

Every awesome AI communicative successful H1 2026 was really a communicative astir attribution.

• We’re still figuring retired really to amended accuracy successful measuring AI visibility.
• Companies spent billions connected inference, but location is nary evident reply to “What’s the ROI?”
• Public markets punished package companies, but were investors reacting to existent aliases perceived disruption?
• We’re proceeding “AI caused layoffs,” but erstwhile you look deeper, the existent causes aren’t connected.
• It’s clear wherever the postulation nonaccomplishment is coming from, but it’s not clear what contented marketplace exemplary will switch it.

The communal thread: AI’s economical effect is expanding faster than we tin property it.

AI Search

In the H1 2025 report, I predicted Google would rotation AI Mode retired further, and that turned retired to beryllium true. AI Mode is now a click distant from AI Overviews, which intends it’s conscionable 2 clicks distant from regular hunt results.

• AI Mode reached 1 cardinal MAU pinch queries astir 3x longer than classical search. 
• Google called it the biggest hunt container upgrade successful 25 years astatine I/O 2026. 
• Gemini 3 auto browse is shipping wrong Chrome. 
• Nick Fox says Google’s AI Mode sends retired billions of clicks to the unfastened web. 

To understand the effect of this shift, I published a bid of personification behaviour studies and information analyses successful the first half of 2026.

The done statement of Growth Memo investigation successful H1 2026:

• Tracking a brand’s beingness successful AI Search is highly complex. You request to facet successful the engine, personalization, reasoning levels, exemplary updates, stochastic variability, etc. Measurement is fragmented. The citation/mention overlap betwixt engines is minuscule: 91% of citations look successful only 1 of ChatGPT, Perplexity, aliases AI Overviews. Prompt search should beryllium person to polling and attraction groups than SEO rank tracking.
• Brand mentions successful AI answers are much impactful connected business outcomes than citations. Don’t get maine wrong, citations style the answer, and for immoderate businesses, that’s astir important. But the mostly of vendors and merchants request to salary attraction to how often they show up successful a sheet of prompts, successful which context, what sentiment, and whether they’re recommended aliases not (ahead of competitors).
• Trust is the highest rate successful AI search. Close to 75% of consumers prime the number 1 consequence successful an AI shortlist. But if they spot a trusted marque anyplace connected that list, they will prime it.
• The mean U.S. big has precocious spot successful AI recommendations. 88% of the time, users accepted AI Mode merchandise recommendations arsenic champion location is. In AI Overviews, connected the different hand, users click, evaluate, and compare a lot. That’s classical hunt behavior, but it doesn’t transportation complete to AI chatbots.
• Writing style, token budget, and method factors matter for agents. If you want to beryllium coming successful AI Search and agentic recommendations, you request to lead pinch unsocial information, constitute successful a direct, easy-to-understand way, region fluff and support your tract technically accelerated and easy to access.

AEO/GEO is simply a marque transmission disguised arsenic a capacity channel. AI recommendations style request (not citations). The portion of optimization is whether AI names, trusts, and recommends your brand.

The grAIpes are sour

November 2025 was a pivotal turning constituent successful AI. Claude’s Opus 4.5 was the first exemplary that was perceived arsenic reliable and robust capable for agentic workflows. A period later, Peter Steinberger went viral pinch Clawdbot, which led to millions of installs, long lines successful China, and Nvidia building a Nemoclaw clone.

Then, things sewage moreover crazier. Shopify, Uber, Meta, and galore different tech giants built token activity boards that incentivized engineers to ray arsenic galore tokens connected occurrence arsenic possible. 

“Let maine springiness you a thought experiment. Let’s opportunity you person a package technologist aliases AI researcher, and you salary them $500,000 a year. At the extremity of the year, I’m going to inquire him really overmuch did you walk successful tokens. And [if] that personification said $5,000, I will spell ape thing else. If that $500,000 technologist did not devour astatine slightest $250,000 worthy of tokens, I americium going to beryllium profoundly alarmed.”

A Meta technologist created an soul leaderboard dubbed “Claudeonomics” that classed complete 85,000 labor by the number of AI tokens they burned. It led to monolithic discarded arsenic labor near AI agents moving connected idle aliases useless tasks conscionable to climb the ranks and gain titles for illustration “Token Legend”.

Reportedly, Meta engineers consumed 73.7 trillion tokens successful conscionable 30 days. (I wonderment really overmuch of OpenAI’s and Anthropic’s maturation was ignited conscionable by workslop.)

Good times. But successful April, CFOs pumped the brakes aft discovering that labor had burned done their yearly token fund successful 4 months. TokenMAXXXING shifted to valueMAXXXING. 

Meta’s leaderboards were unopen down successful April (along pinch leaderboards astatine different companies), but not earlier the #1 personification connected the committee averaged 281 cardinal tokens, which would costs complete $1.4 cardinal astatine modular API rates.

But what remains from the all-you-can-eat token craze is simply a measurement alteration successful productivity that we cannot yet measure. As I wrote, AI changed my work. And yours, too:

I person utilized AI to raise my effect by a magnitude.

• I designed landing pages end-to-end for a awesome recreation marque that made it into production.
• I automated taxable prioritization, SEO testing, and SEO reporting for my clients pinch full-blown apps.
• I built an array of useful applications for myself, from automating the SSI (SEO Site Index recovered successful the bimonthly Growth Intelligence Briefs) to Openclaw agents that thief maine pinch investigation and charts.

One result of the token tsunami is that a batch much group started utilizing Claude, which successful return fired up the maturation engines for AI infrastructure companies that Claude recommended, for illustration Supabase.

Source: Supabase Series F

SaaS bloodbath

In February 2026, arsenic a consequence of the Opus 4.6 merchandise and Claude Cowork, the software-as-a-service manufacture saw a terrible marketplace downturn that wiped retired hundreds of billions of dollars successful marketplace headdress complete conscionable a fewer trading sessions. The SaaS vertical saw a steep yearly diminution of +30% successful the markets. Valuations for high-growth package companies collapsed, pinch gross multiples shrinking dramatically compared to their pandemic highs.

If you’re moving successful B2B, you’ve astir apt felt the effect successful the shape of little conversion rates, longer income cycles, and softening marque searches.

The Brand Tax Calculator gives you the dollar fig for the spread betwixt branded paid walk and reported ROAS, and the AI SEO Budget Reallocation Planner turns it into 5 growth-priority scenarios successful nether 7 minutes. Both unrecorded successful the Premium Resource Library. 

Source: It’s Time to Build, July 17, 2026 

It turns retired not each package was (or is) forsaken. It’s the bottommost quartile of package companies that pulls the marketplace down. Interestingly, the diminution is purely associated pinch really “the market” views a company’s AI robustness. Company valuation drops are not related to performance, but alternatively to whether group deliberation AI will disrupt their product.

Over the past 30 trading days, some the apical quartile and median IGV banal person outperformed the ETF arsenic a full — it’s the bottommost quartile (which see immoderate of the largest companies) that’s pulling down wide performance.

So, if you’re successful B2B and either forced aliases privileged to return a caller job, prime a institution the marketplace is optimistic about.

AI washing successful the labour market

You mightiness person publication aliases heard this recently: “We’re laying disconnected x% of our workforce owed to AI.” It’s the narrative du jour. Challenger, Gray & Christmas reported that AI was the starring cited logic for May 2026 occupation cuts and had been cited successful 87,714 cuts year-to-date, adjacent to 22% of each 2026 announced layoffs done May.

Source: May 2026 Challenger Report

And don’t get maine wrong. Layoffs are real:

• Tech layoffs up ~66% twelvemonth complete year, toward 150K
• Oracle trim 21,000 citing AI
• Challenger lists AI arsenic the apical cited logic (~87,700 YTD by May)
• Block trim 40% and its banal roseate (Bloomberg’s “AI washing” framing)

The only problem is that AI is not the cause. I’ve written extensively astir the Big Labs utilizing AI replacement arsenic a trading narrative. But it hasn’t travel existent yet. A batch of AI layoffs were really to offset superior expenditures, pandemic overhiring, and economical turbulence. We tin each ideate a early successful which AI is truthful bully that it makes definite roles obsolete, but that clip is not now.

A twelvemonth ago, successful the H1 2025 report, I predicted that AI layoffs were a PR stunt. And I was right: Despite clear productivity gains from AI, the caller waves of layoffs show no indication of AI being the underlying cause. In fact, companies utilizing AI arsenic a logic to laic labor disconnected are rehiring them.

The supplier marketplace fragmented

ChatGPT’s marketplace stock decreased from 78% successful July 2025 to 56% successful July 2026, while Gemini went from 15% to 30% and Claude from 2% to 10%.

Google is now the astir apt subordinate to triumph the AI user market, and ChatGPT has mislaid its value successful AI Search to Google:

• ChatGPT still has ~1.1B users, but OpenAI refocused connected an endeavor pivot and unopen down Sora, dropped video successful ChatGPT and killed Instant Checkout. Enterprise makes up ~40% of revenue, heading to ~50% up of a imaginable IPO. I’m skeptical that we’ll beryllium capable to bargain OpenAI stock. According to Ed Zitron, OpenAI spends $2.8 cardinal USD each period to make $1.1B.
• Anthropic’s conflict pinch the Pentagon pushed Claude to No. 1 among free apps connected Apple’s U.S. App Store. A fewer months later, users look little judge if Anthropic tin defy attacks from open-source models.
• The Big Labs massively subsidize tokens. SemiAnalysis recovered a $200 USD scheme of the frontier models gives you 8-14K USD worthy successful tokens. 

Lately, open-source models for illustration Kimi K3, GLM 5.2, aliases Deepseek V4 person put unit connected the large US labs because of those subsidies. Open-weight and open-source models besides show that the exemplary itself becomes commoditized, while the harness and exertion furniture summation much importance. 

Forward-deployed-engineer job postings roseate 800% successful the first 9 months of 2025 and were still up 729% twelvemonth complete year successful April 2026. Frontier Labs now pays more than $500,000 successful full compensation for group who tin move a exemplary demo into a moving strategy wrong a customer’s business.

Never person marketers faced a faster evolving situation pinch much platforms to measure… which is why knowing what platforms matter astir for your assemblage is important. But from an execution perspective, exemplary readiness has go a business risk, arsenic we tin spot astatine the manus of Fable 5. Having options is good.

Publishers and the law

I predicted up to 70% of 2024 integrated postulation could beryllium gone by 2026, but that was overeager. In the end, the numbers onshore astatine ~33% referral nonaccomplishment twelvemonth complete year, pinch 68% of Google searches now ending without a click. Still, arsenic a result, publishers are moving the conflict to the courts and regulators.

USA Today CEO Mike Reed explained: “We’re getting adjacent now to the constituent wherever we’ll artifact Google arsenic good and wantonness the accepted hunt postulation that we get today.” 

Here are conscionable a fewer lawsuit-related highlights:

• A tribunal successful Munich ruled Google liable for mendacious statements generated by AI Overviews (June 2026).
• 400 newspapers sued OpenAI and Microsoft complete unauthorized contented usage (June 2026).
• The UK CMA ordered Google to springiness publishers greater power and transparency complete really contented is utilized successful AI hunt (including opt-outs from AI Overviews, AI Mode, and Discover summaries), on pinch clearer attribution and improved reporting connected really contented appears successful AI-generated results.⁠ Google responded by adding an impression-based AI study successful Search Console.

In consequence to the aggravated publishing scenery changes, Parallel AI and Cloudflare are some trying to build publishing marketplaces for contented owners and AI companies. 

“Assume there’s nary search. You person to person your businesses planned arsenic if hunt is zero.”

Publishers are trying to switch the content-for-traffic bargain pinch a content-for-training market. This leaves respective unfastened questions:

• Who sets the value and how? 
• How tin you beryllium that an article contributed to an AI agent’s reply aliases purchase? 
• Do publishers person speech power? 
• Are smaller publishers moreover needed by AI labs?

Dig in: Full database of Growth Memo investigation completed successful H1 2026

Dig deeper into the patterns covered successful this halftime study below:

• User behaviour studies:
  • How consumers navigate high-stakes purchases successful AI Mode
  • Users behave otherwise successful AI Overviews vs. AI Mode
  • What to do now that AIOs turned hunt into reference sessions
• Data analyses:
  • The Consensus Gap
  • The shade citation problem
  • Reasoning lift: What happens to AI visibility erstwhile AI thinks harder
  • The subject of really AI pays attention
  • The subject of really AI picks its sources
  • The subject of what AI really rewards
  • Shorter, Focused Content Wins successful ChatGPT
  • GSC information is 75% incomplete
  • Organic rankings vs. merchandise grids: The caller e-commerce divide
  • The AI skills net premium
  • Where AI agents get stuck connected your site
  • Why astir original information ne'er gets cited

As we walk the halfway constituent successful the year, a large acknowledgment to my editorial and investigation partners, Amanda Johnson and Eric van Buskirk. And an hold of gratitude to my information partners who make these analyses possible: Semrush, AirOps, Gauge, Citation Labs, Siteline, SalaryGuide, and AudienceKey. 

This station first appeared connected the author’s website and is republished present pinch permission.