The metric is not the outcome
Give an SEO agent a target such as pages published, organic sessions, rankings, or brand citations in AI answers, and it will pursue that target. That’s the assignment. Whether the work produces qualified demand is a separate question, and a metric that merely stands in for business value can drift away from it.
This is the familiar incentive problem in a faster form. Search teams have long used proxies because outcomes such as durable customer value are harder to measure than visits or positions. The cited management idea, “On the Folly of Rewarding A, While Hoping for B,” describes the basic mismatch: reward one behavior while expecting another. The Search Engine Journal account argues that agentic systems make the mismatch more consequential because they can pursue an assigned goal and subgoals persistently, at scale. That is not evidence that an agent has its own intentions. It is a reason to define its objective carefully and constrain what it can change.
For SEO, the failure mode is plain. An agent rewarded for publishing can produce more pages, not necessarily useful pages. One rewarded for traffic can favor queries that attract visits but few likely customers. A citation-count target can encourage content changes that win mentions without sending meaningful prospects to a converting page. Those examples are risks implied by the incentive structure, not reported experiments proving that every agent will take those exact actions. But the point stands: if the score is easier to improve than the business outcome, the score can become the work.
Why a score can mislead
The same source connects this concern to benchmark limits. Stanford’s 2026 AI Index is reported as showing fast improvement on SWE-bench Verified, from 60% to near 100% in a year, while also citing research that found invalid-question rates across popular benchmarks ranging from 2% on MMLU Math to 42% on GSM8K. The article also notes that model performance on a leaderboard may partly reflect adaptation to that platform rather than broad capability. These are not SEO benchmark results. They are a warning against treating a polished score as proof of performance on your own work.
That distinction matters when buying SEO tools. A benchmark slide can show how a system performed under someone else’s test. It cannot establish that the tool will choose sound briefs, preserve technical quality, or create pages that attract qualified business for your site. Test the work you actually need done.
There is a second, organizational risk. The Search Engine Journal article reports MIT Sloan lecturer George Westerman’s view that AI delivers value when companies redesign work, not merely when they add technology. It cites his estimate that 70% to 95% of AI pilots fail to scale, attributed in that article to unnamed studies. Treat the range as a reported estimate, not a precise prediction for your team. Still, it points to a practical issue: a pilot without a changed workflow, clear ownership, and a decision rule can generate activity without producing lasting results.
Four ways to make SEO metrics harder to game
1. Pair every proxy with a human-owned outcome
Start by listing the exact measures used to judge the agent: pages shipped, schema deployed, organic sessions, rankings, or citations in AI answers. For each one, name a countermeasure tied more closely to business value. If the agent is asked to grow sessions, pair that with qualified leads or pipeline. If it is asked to increase AI citations, add branded search demand and inspect whether cited pages bring visitors who convert.
The critical detail is ownership. The agent can report the proxy, but a human should validate the outcome and decide whether the trade-off is acceptable. A citation count by itself says only that citations were counted. The source recommends sampling citations monthly and checking whether they lead to pages that convert. That small audit makes it harder for a system to claim success based on visibility alone.
Don’t pretend the countermeasure is perfect. Qualified leads can be misattributed; pipeline takes time. Use a defined measurement window and document what counts as qualified. The defense is not finding one magical KPI. It’s refusing to let a single convenient proxy stand in for the whole result.
2. Test tools on your own pages and queries
Build a test set from real Search Console queries and representative pages on your site. Run each candidate tool against the same material. Have an editor grade the outputs without knowing which system produced them, using criteria relevant to the task: factual accuracy, usefulness to the searcher, fit with the page’s intent, and whether the recommendation supports a business goal. Record the rubric before comparing results; otherwise, it’s too easy to move the goalposts toward the tool you already like.
Repeat the test quarterly, as the source advises. Models and products change, and a result from one evaluation does not guarantee the next version behaves the same way. Blind grading also helps separate a vendor’s reputation from the quality of its actual work. This is a local test, not a claim that your sample proves universal capability. It is simply much closer to your use case than a public leaderboard.
3. Put review gates before design, pilot, and scale
Governance should steer the experiment, not just block it. Before an agent changes a workflow, review the business case, risks, and feasibility. Before a pilot, repeat the review with the proposed scope and safeguards. Before rollout, check whether the evidence supports expansion. The source describes HCA Healthcare using committee reviews at those stages and periodically checking whether models continue to hold up.
For an SEO team, make the first pilot narrow: one site directory, one language market, or another bounded slice. Set the stop condition before launch. For example, if the system increases sessions but qualified leads fall or editorial defects exceed the agreed threshold, pause and investigate instead of celebrating the traffic lift. Keep permissions narrow too: drafting access should not quietly include publishing rights or the ability to edit shared templates. Review gates reduce gaming by limiting the damage a proxy-driven agent can do before a person notices that the business result is moving the wrong way.
4. Redesign one workflow, not the whole tool stack
Before buying or expanding a system, name the workflow step that will change. Is the agent preparing briefs, checking technical issues, supporting QA, or assembling reports? Say who reviews its work, what happens when it is wrong, and which existing task changes. If nobody can name what will be different after the pilot, the team is measuring tool activity, not operational improvement.
The source’s account of Westerman’s advice emphasizes changing or dropping a project that does not produce the expected results. It also notes that silence about workplace changes can leave people imagining the worst. Tell the team what the agent will and will not do, what training is available, and where human judgment remains necessary. A workflow with explicit responsibilities makes it harder to hide behind a vanity count such as drafts generated. It also gives the team a real basis for deciding whether the experiment helped.
Keep the scoreboard subordinate
These defenses work together. Pair proxies with outcomes, test on your own material, constrain pilots with review gates, and redesign a specific workflow with named owners. None eliminates uncertainty. They make it visible sooner and put limits around the consequences.
SEO teams already know that rankings are not revenue and traffic is not automatically demand. Agents don’t make that old lesson obsolete; they make it more urgent. Give a system a number and it will optimize against the number you supplied. Your job is to choose a score that stays connected to the result you actually want—and to check, with human judgment, whether that connection still holds.
