Profoundly Wrong (Part 2): Why AI Marketing Isn't Working

Industrial content production carries costs that a citation dashboard does not show

The market's answer to AI search has been production at industrial scale: more pages, more agents, more domains, and engineers embedded in client teams to push it through. In AIVO's multi-turn testing, the volume that wins citations does not decide which brand an AI system recommends. Independent research finds no optimization technique with a stable effect on discovery or traffic, and some that make retrieval worse.

Key findings. Across more than 25,000 multi-turn probes covering over 250 brands, AI systems displaced a brand they had already cited in 87.3% of cases before recommending. In repeated identical runs, the sources cited changed 95.8% of the time while the chosen brand changed 14.6% of the time. A July 2026 review of 45 studies found that optimising only a page's body cut its top-10 retrieval presence by 16%.

The industrial answer

As AI assistants began answering buying questions, marketing reached for the tools it knew. If AI systems cite content, produce more content. If they cite it unevenly, produce it faster, in more formats, across more domains.

The most ambitious version of this now runs at industrial strength. Agents draft pages by the thousand. Engineers are deployed inside client teams to wire production into publishing systems and push volume through. Dashboards track the citations and visibility that result.

This response assumes that AI recommendation is primarily a retrieval problem. Our findings suggest it is primarily a decision problem.

Why volume misses the decision

AIVO's research points to a different diagnosis. In multi-turn conversations, brands rarely lose because the AI system cannot find them. They lose later, when it compares options against the customer's criteria and decides.

In working paper WP-2026-14, across more than 25,000 multi-turn probes covering over 250 brands, AI systems displaced a brand they had already cited in 87.3% of cases before reaching a recommendation. Those brands had cleared the supply test. Their content was retrieved and used. They were still not chosen. Retrieval is necessary. It was not sufficient.

Our citation stability work shows why counting citations hides this. In repeated identical runs, the sources an AI system cited changed 95.8% of the time, while the brand it finally chose changed 14.6% of the time. Content production moves the volatile layer. The decision sits in the stable one. Our reading is that it turns on whether the AI system can justify choosing the brand against the customer's criteria, which page count alone does not supply.

What independent research finds

Our findings are consistent with several independent studies.

A critical survey of 45 generative engine optimisation studies, posted to arXiv in July 2026 by Olivier Martinez of Sciences Po, concludes that no reviewed technique produces a stable, cross-platform effect on discoverability or downstream traffic. In one end-to-end test across 171,003 documents, optimising only a page's body reduced its top-10 presence after reranking by 16% and its final citation by 6%. In another benchmark, only 3 of 54 method and domain combinations were significantly positive.

The widely quoted claim that optimisation lifts visibility by 40% traces, according to the survey, to one metric in one configuration, where the documents had already been handed to the model.

Practitioner data points the same way. A 2026 study by Leoprd found that brands which appeared in AI answers but lost to a competitor averaged more citations, 9.2, than brands chosen as the primary recommendation, at 8.4. A March 2026 analysis by Seer Interactive of 541,213 AI responses found content being cited while competitors were recommended, and concluded that adding more content will not fix that pattern and may make it worse.

The costs nobody prices

Industrial content carries costs that a citation dashboard does not show.

The first is the brand. A multi-turn buying conversation is an interrogation rather than a search, and every recommendation invites the next question: why this brand? An AI system answering it looks for reasons it can verify and repeat. A thousand pages that restate the same claims give it more to cite and nothing new to defend.

The second is retrieval itself. The survey's end-to-end results show that a page rewritten to perform better once selected can become less likely to be selected at all. Volume produced to a formula can lower the odds it was meant to raise.

The third is the line between optimisation and manipulation. Both work by changing the text an AI system reads, and the survey proposes truthfulness, verifiable evidence and disclosure as the tests that separate them. Content produced at volume is hardest to hold to those tests, and practitioners warn that penalties for aggressive tactics now carry across both traditional and AI search.

What we are not claiming

We are not claiming that content does not matter. A brand an AI system cannot retrieve cannot be recommended, and controlled experiments show that information placed in an AI system's working context can influence its outputs and recommendations. The same survey finds query relevance the most reproducible lever, with verifiable, dated evidence close behind.

Our claim is narrower. In multi-turn conversations, the citations a brand earned did not predict whether it was chosen at the point of decision. Content produced to earn citations therefore buys a measure that moves, while the decision stays where it was.

What to measure instead

If the decision is where brands win or lose, the decision is what to measure. That means testing whether AI systems choose the brand across full conversations, against the criteria real customers bring, and repeating the test often enough to separate change from noise.

It also changes the brief for content. The useful question for each page is whether it gives an AI system a reason to choose the brand that the system can verify and repeat. A smaller body of specific, evidenced material may do more than a large body of restatement.

The test for any approach, including ours, is whether it changes recommendation behaviour rather than visibility metrics.