AI Agents Will Game Your SEO Metrics, MIT & Stanford Research Points To The Risk

In case your search engine optimisation crew is handing extra of its work to AI brokers, the metric you reward these brokers for will matter greater than the mannequin you select. Current items from MIT and Stanford, learn side-by-side, level to that conclusion, and every comes with a repair you possibly can drop into subsequent quarter’s plan.

The primary is an interview that Joshua Miller of The Boston Globe ran in his Camberville e-newsletter on September 17. His visitor was Dylan Hadfield-Menell, an affiliate professor {of electrical} engineering and pc science at MIT on the school of synthetic intelligence and decision-making. Hadfield-Menell research how objectives get set for AI programs and the way that course of goes unsuitable. He opened with an instance each search engine optimisation will acknowledge.

The Vacuum That Fed Itself

Researchers as soon as educated a robotic vacuum with reinforcement studying, rewarding it each time it picked up filth. The vacuum discovered to select up filth, dump it again on the ground, and decide it up once more. It hit the goal and defeated the aim.

Hadfield-Menell connects that story to a Nineteen Seventies administration paper titled “On the Folly of Rewarding A, While Hoping for B.” Its traditional case is the college professor who will get promoted for publishing analysis whereas being anticipated to show. Pay for one conduct, and also you get that conduct, no matter you had been hoping for.

What has modified, he says, is scale. Since early 2025, builders have utilized reinforcement studying at a lot bigger quantity on prime of language fashions, and it strengthens some behaviors no person needs. He pointed to a latest incident involving OpenAI programs and Hugging Face, the place fashions that judged a activity too arduous went on the lookout for methods to cheat the check. He in contrast it to breaking right into a professor’s workplace to steal the examination.

His fear will not be that machines get up with objectives of their very own. Techniques are handed a objective, undertake subgoals alongside the best way, and preserve pushing towards completion in a means he referred to as “sticky.”

My view is that search engine optimisation is the occupation greatest positioned to know this downside and the slowest to confess it applies to us. We now have spent greater than 20 years optimizing proxies. Rankings, site visitors, area scores, and now AI visibility scores all stand in for a enterprise end result that no person can measure straight. A human crew video games a proxy slowly and with some hesitation. An agent does it quicker and with none.

The Scoreboard Is Shakier Than Distributors Admit

Stanford’s 2026 AI Index exhibits why leaning on printed scores is dangerous. The report says AI retains enhancing rapidly. On SWE-bench Verified, a coding benchmark, efficiency rose from 60% to close 100% in a single yr, and 88% of organizations now use AI.

The identical report, in its technical efficiency chapter, cites a evaluate that discovered invalid-question charges on in style benchmarks starting from 2% on MMLU Math to 42% on GSM8K. It additionally notes analysis suggesting {that a} mannequin’s standing on the Enviornment leaderboard might partly mirror adaptation to the platform relatively than basic functionality.

Michelle Kim of MIT Technology Review summarized the report in April. She provides that fashions educated on benchmark check information can be taught to attain effectively with out getting smarter, and that the highest fashions now sit very shut collectively and compete on price, reliability, and real-world usefulness. Yolanda Gil, a College of Southern California pc scientist who coauthored the report, informed Kim that when an organization leaves out its outcomes on sure benchmarks, significantly the responsible-AI ones, the omission “possibly says one thing.”

That ought to change how an search engine optimisation crew retailers for instruments. If the main fashions sit inside a couple of factors of one another, and the scores themselves could be flawed or gamed, a vendor’s benchmark slide tells you little about how the product will deal with your pages, your queries, and your shoppers. I’d belief one check alone website over any leaderboard.

See additionally: The 4-Step Test That Catches AI Errors Before They Shape Your Strategy

The place The Returns Really Come From

MIT Sloan’s Betsy Vereckey reported in August on the query that follows, which is what separates companies that profit from AI from those that don’t. George Westerman, a senior lecturer at MIT Sloan and a digital fellow on the MIT Initiative on the Digital Financial system, says the reply will not be higher algorithms. The winners redesigned how work will get performed. On the MIT Enterprise AI Discussion board in Could, he informed the viewers that know-how delivers little till the enterprise itself operates in another way.

He additionally put the share of AI pilots that by no means scale at someplace between 70% and 95%, a spread the Sloan article attributes to research with out naming them. Pilots are straightforward to launch and arduous to unfold.

Westerman’s sharpest check for leaders is about governance. “Is your governance extra the steering wheel or is it extra the brakes?” he requested. HCA Healthcare exhibits what the steering-wheel model seems like. A committee opinions the dangers, enterprise case, and feasibility of each AI use case, then asks its questions once more earlier than a pilot at a small variety of hospitals and once more earlier than the venture scales. It additionally checks periodically that its fashions are nonetheless holding up. The chance questions level the crew towards what to research relatively than stopping the work.

Advertising and marketing seems in Westerman’s case research too. Dentsu Inventive has pushed AI throughout planning, inventive, market analysis, and marketing campaign work.

I believe a number of search engine optimisation groups working AI pilots are headed for that 70% to 95%. A pilot that doesn’t change the temporary, the evaluate step, or the reporting is a software trial, regardless of the slide deck calls it.

See additionally: Why 88% Of Companies Are Using AI Wrong: The System-Building Gap

How To Apply This To Your search engine optimisation Technique

4 strikes comply with from the three items.

Pair each proxy with an consequence the agent can’t contact. Checklist the metrics your AI-assisted workflows are judged on, from pages printed to schema deployed to model mentions in AI solutions. Then connect a second measure {that a} human owns, reminiscent of qualified leads, pipeline, or branded search demand. I like Citation Share of Voice, however it’s a proxy. If a content material agent is judged on how typically your model exhibits up in AI solutions, anticipate it to seek out the most affordable route there. Verify a pattern of these citations by hand every month and see whether or not they ship anybody to a web page that converts.

Take a look at instruments by yourself pages. Pull a set of actual queries from Search Console, run every candidate software in opposition to your individual content material, and have an editor grade the outcomes with out figuring out which software produced them. Repeat it each quarter, as a result of the leaderboards will transfer and also you received’t know why.

Gate the brokers the best way HCA gates use instances. Add evaluate factors earlier than design, earlier than the pilot, and earlier than scale. Pilot on one listing or one language market, and resolve prematurely what end result ends the pilot. I’d additionally preserve agent permissions slender, so drafting doesn’t quietly flip into publishing or enhancing templates. Westerman’s recommendation is to alter or drop a venture that isn’t producing the outcomes you anticipated.

Rewrite one workflow, not the software stack. Earlier than shopping for something, identify the step in your course of that shall be completely different after the pilot, whether or not that’s briefing, QA, or reporting, and say who loses a activity due to it. Then inform the crew what adjustments and what coaching comes with it. Westerman notes that silence lets individuals think about the worst.

I don’t suppose the following mannequin launch will resolve who wins in AI search. The metric you utilize to guage your brokers will, as a result of the brokers will discover it earlier than you do. Select one you’d be glad to see them hit.

Extra Assets:

Featured Picture: Fardived/Shutterstock