ED ZITRON: AI is not making businesses outside the AI sellers better, and at Google it is weakening the core search business

The Gist

Ed Zitron says that outside the companies selling AI, businesses are not getting more productive or profitable from it. Google even said its AI search makes less money than normal search, and it is spending about $50 billion a quarter to build it. Four years in, he thinks the payoff should show by now. This is a steelman: Ed Zitron's arguments from the conversation are reconstructed in their strongest fair form, with his answers to the host built in. It does not endorse them.

Conclusion

So far the evidence shows AI paying off mainly for the companies that sell it, while at Google it is slowing the core business it was meant to strengthen.

Premises

1. Apollo's chief economist Torsten Sløk finds revenue per worker and profit margins flat for the S&P 493 and no sign of AI lifting productivity, while mainstream companies have begun treating token spending as a cost.
2. Google said the growth of AI Mode queries hurt search because AI Mode monetizes at a lower rate; Alphabet's services revenue missed estimates and its shares had their worst day on record.
3. Google spent about $51 billion on capital expenditure in the quarter, largely to build that AI capacity.
4. Amazon missed estimates on online stores and advertising, spent $77 billion on capital expenditure in the quarter, and plans more in 2027; only its cloud unit beat.
5. After four years and with a majority of venture capital going to AI, the payoff should be visible by now ("time compression").

Assumptions

• Disclosure: Ed Zitron is CEO of EZPR, a media relations and public relations firm. He writes the Where's Your Ed At newsletter, which has a paid premium tier, and hosts the Better Offline podcast, produced with iHeartRadio and Cool Zone Media; both cover the AI industry critically, and in this conversation he says he predicted the rating outcome ("told you so"). Sources: "About Where's Your Ed At" (Where's Your Ed At) at https://www.wheresyoured.at/about/, "Introducing: Better Offline" (Where's Your Ed At) at https://www.wheresyoured.at/introducing-better-offline/ and "Ed Zitron" author page…
• The conversation was recorded on October 30, 2026, as the FT story on Anthropic's expected rating broke; figures are as stated live and may be approximate.
• "S&P 493" means the S&P 500 excluding the seven largest tech companies.
• Differs: Alphabet revenue still rose 25 percent and search revenue about 12 percent, while Google Cloud grew 103 percent and beat estimates.
• Differs: AWS grew 39 percent and beat the consensus of 38 percent, and the Prime Day shift into Q2 hurt Q3 stores and ads; the "Deutsche Bank 42 percent" figure was not found.

Analysis

Argument type: Inductive.

Potential Fallacies

• Cherry-picking / selective emphasis (P2, P3, P4 and the conclusion) — Misses, capex totals and the record stock drop are foregrounded, while 25% revenue growth, 12% search growth, 103% Cloud growth, the AWS beat and the Prime Day timing shift are treated only as caveats. The headline therefore reads more negatively than the full data. The Cloud strength also sits awkwardly with 'AI is hurting Google' overall.
• Questionable cause / correlation as causation (P1, P2, P4 to conclusion) — Google's statement supports a narrow causal link (lower monetization per AI Mode query), but not that AI is net-weakening the core business. Amazon's retail and ads miss is placed beside AI capex although the Prime Day shift offers a concrete non-AI explanation. Flat S&P 493 margins are likewise read as AI's failure without ruling out tariffs, rates, labor costs, and adoption lag.
• Spending treated as evidence of failure (P3 and P4) — Capex is an input and a forward-looking investment. Large figures show commitment and expected demand, not poor returns. Without return, backlog, or utilization data, P3 and P4's capex numbers add little to the case.
• Arbitrary timeline / argument from absence of evidence (P5 and P1) — 'Payoff should be visible by now' sets a deadline without a benchmark. Earlier general-purpose technologies (electricity, IT, ERP) took years or decades to appear in aggregate productivity data, and four years from ChatGPT's launch is only a few years of serious enterprise deployment. Treating a missing aggregate signal as proof of no effect is not secure, and the claim is hard to falsify.
• Ecological inference / aggregation (P1) — Flat averages across roughly 493 firms can hide heterogeneity, such as some adopters gaining and others losing. They also miss benefits that don't show up in revenue per worker or margins (quality, avoided hiring). Counterfactual margins might have been lower without AI.
• Hasty generalization / unrepresentative examples (P2–P4 to conclusion) — Single-quarter results at two firms and one economist's aggregate stand in for 'businesses outside the AI sellers'. Google and Amazon are themselves major AI sellers, so they are weak examples of the non-seller category.
• Unfalsifiable framing and unestablished contrast term (Conclusion) — Seller strength (Cloud, AWS) is absorbed as confirming the thesis, so favorable results never count against it. Revealed demand from paying customers also cuts against 'no one else benefits'. The claim that AI pays off 'mainly for sellers' offers no direct data on seller profitability.
• Market reaction treated as business performance (P2) — Estimate misses and 'worst day on record' measure deviation from expectations, capex reaction, and guidance, not the underlying health of search. For a mega-cap, a record move also partly reflects scale and elevated expectations.

Counterarguments

These challenges are not ranked.

• Conclusion (Google half) (High impact) — AI Mode is likely defensive: without it, search might lose share to chatbots and other answer engines. Lower per-query monetization is then the cost of staying relevant, and the relevant comparison is AI versus no AI, not AI versus the old status quo. With search revenue up about 12% and total revenue up 25%, 'weakening' overstates slower or lower-yield growth. Monetization may also recover as ad formats adapt to AI answers.
• Premise 1 (High impact) — Aggregate S&P 493 revenue per worker and margins are a coarse, confounded test. Tariffs, rates, wages and sector mix all move them, and the technology's productivity effects typically lag adoption (the electricity and IT productivity paradoxes). Firm-level and use-case evidence (coding, customer support) may show gains that averages hide, and benefits like cost avoidance and quality do not appear in these metrics.
• Conclusion (sellers half) and Premise 4 (Medium impact) — Cloud growth of 103% at Google and 39% at AWS reflects customers paying real money for AI capacity, which is revealed demand and sits uneasily with the claim that no one outside the sellers benefits. Google and Amazon are also both sellers and users, which blurs the categories. Alternatively, if that demand is itself circular or speculative, the sellers' payoff may be less secure than the argument concedes.
• Premises 3 and 4 (Medium impact) — Capex is a multi-year investment recouped through cloud backlog, not a quarterly outcome. Comparing one quarter's spend against one quarter's results mismatches time horizons, and the Prime Day shift explains Amazon's stores and ads miss without reference to AI.
• Premise 5 (High impact) — Four years since ChatGPT is only a couple of years of serious enterprise deployment. Prior platform shifts (ERP, cloud, the internet) took longer to show up in aggregate data and often depressed margins first. Few serious AI proponents claim transformation by now, so the premise attacks a weak version of the opposing view.
• Conclusion (source reliability) (Low impact) — The case comes from a single openly skeptical commentator with commercial and reputational stakes who claims to have predicted the outcome, speaking live from memory (one cited figure, 'Deutsche Bank 42 percent', could not be verified). This does not falsify the confirmed facts but warrants independent corroboration of the framing and a check on selective emphasis.

Suggested Improvements

• Separate the two claims and scope them properly — Restate the conclusion as: 'So far AI revenue gains are concentrated in infrastructure sellers, and at Google lower AI Mode monetization is a measurable drag on search growth.' Drop or substantiate 'weakening' and 'slowing the core'. The narrower claim is supported by the evidence, while the broader one conflicts with the conceded growth figures.
• Add a counterfactual for Google — Provide data on AI Mode revenue per query versus classic search, the net effect on total Google revenue, and evidence about what share Google would have lost without AI Mode. Without a no-AI-Mode baseline, a monetization drag cannot be distinguished from a defensive trade-off.
• Use firm-level evidence on adopters — Compare AI adopters with matched non-adopters on revenue per worker and margins, with 2–5 year lags, or cite enterprise ROI surveys and controlled studies. Check the Apollo result across alternative cohorts and time windows. This addresses the ecological and confounding problems with aggregate S&P 493 data.
• Justify the timeline with a reference class — Compare against the diffusion lags of electricity, PCs, ERP, and cloud, and pre-specify what magnitude and horizon of payoff would falsify the 'no payoff' thesis. This turns 'time compression' from an assertion into a testable claim.
• Give return metrics alongside capex — Report capex relative to revenue, cloud backlog, utilization, depreciation schedules, and return on invested capital, and separate AI from non-AI capex. Capex figures only matter when set against returns.
• Integrate the counter-facts and adjust for timing — Treat Cloud +103%, search +12%, the AWS beat, and the Prime Day shift as evidence to weigh in the conclusion, for example by comparing Amazon's Q2+Q3 combined year over year. Corroborate Apollo and other figures from primary sources. This reduces the appearance of selectivity and makes the argument robust to obvious rebuttals.

Scenario Tests

• Over the next several quarters, AI Mode monetization catches up as ad formats mature and Google's search growth holds or re-accelerates. (Challenges) — The Google half of the conclusion would look like a transitional mix effect, and the single-quarter evidence would not support 'weakening the core'.
• Later Apollo data or firm-level studies show productivity and margin gains concentrated in AI-adopting companies, appearing with a lag. (Challenges) — The claim that AI is 'not making businesses outside the sellers better' would fail. Flat aggregates would be revealed as a lagging or masked signal.
• Combined Q2+Q3 Amazon results, adjusted for Prime Day timing, show stores and ads roughly on trend. (Challenges) — P4 would lose most of its probative value and leave only the capex totals, which show spending, not failure.
• AI-seller revenue turns out to depend heavily on circular financing among labs, clouds and chipmakers, and customer spending contracts as firms treat tokens as a cost to be capped. (Supports) — The 'payoff mainly for sellers' claim would be weakened in one direction (sellers' gains are fragile) but the broader skeptical thesis that customers are not realizing value would gain support.
• Google's search revenue per query falls further and its total search revenue growth turns flat or negative while AI Mode share keeps rising, with no offsetting competitive-share benefit. (Supports) — This would substantiate cannibalization as a net harm to the core business, not just a slower growth rate.
• Hyperscalers continue to report supply-constrained cloud demand with rising backlog while customer ROI evidence remains absent. (Neutral) — This is consistent with both 'sellers profit while buyers lag' and 'early-stage adoption with returns to come', and does not discriminate between them.

Coherence & Relevance

The argument is a cumulative-case induction whose individual facts are mostly confirmed but whose inferential links are thin. The sellers-benefit half is coherent and plausible, although it relies on cloud growth rather than direct seller-profit evidence and is hard to falsify. The Google-harm half overreaches: the author's own conceded figures (search +12%, Alphabet +25%, Cloud +103%, AWS beat, Prime Day shift) point to a more mixed picture of lower-yield growth during a transition. P3 and P4 supply context, not support, and P5 does inferential work without being evidence. The conclusion's hedges ('so far', 'mainly') soften it but do not close the gaps in causation, counterfactual, or timeline.

• Apollo's chief economist Torsten Sløk finds revenue per worker and profit margins flat for the S&P 493 and no sign of AI lifting productivity, while mainstream companies have begun treating token spending as a cost. (Strong) — Directly relevant to the 'no benefit outside sellers' claim, but aggregate flatness does not isolate AI's effect, ignores adoption lag and heterogeneity, and the S&P 493 includes some AI sellers and enablers. The token-as-cost claim is unsourced.
• Google said the growth of AI Mode queries hurt search because AI Mode monetizes at a lower rate; Alphabet's services revenue missed estimates and its shares had their worst day on record. (Strong) — Supports a narrow monetization headwind but not net weakening. The argument does not engage with the 12% search growth or 103% Cloud growth, the defensive counterfactual, or the fact that the stock move reflects expectations and capex.
• Google spent about $51 billion on capital expenditure in the quarter, largely to build that AI capacity. (Weak) — Capex measures input, not outcome. It is equally compatible with demand-driven buildout, and no return, backlog, or utilization data link it to a payoff failure.
• Amazon missed estimates on online stores and advertising, spent $77 billion on capital expenditure in the quarter, and plans more in 2027; only its cloud unit beat. (Weak) — The retail and ads miss has a non-AI explanation (Prime Day timing), the AWS beat is marginal, and Amazon is itself an AI seller. The premise implies a pattern with Google without a causal link.
• After four years and with a majority of venture capital going to AI, the payoff should be visible by now ("time compression"). (Moderate) — It supplies the missing bridge from 'no visible gain' to 'no gain', but it is an unsupported normative expectation without a benchmark from comparable technology cycles.

View this argument on LogicFirst.ai