What not to do

The cheapest improvement in most AI visibility budgets is to stop doing things that do not work. This page is that list: the tactics that are either measurably ineffective in the controlled literature or explicitly disclaimed by the search engines that run the answer surfaces, each with its source and date. It names tactics, not firms, because a firm can change what it sells and a tactic either works or it does not. Each item below states who said it does not work, when, and where to read it yourself. The list is short on purpose: it covers only what has been measured or disclaimed, not what we happen to dislike. The scorecard at the end turns Google's own vendor-vetting questions into a checklist you can take to whoever currently sends you invoices, and it prints.

The engines have said what they do not read

On 15 May 2026 Google Search Central published a guide to optimising for generative AI features in Google Search, updated 10 July 2026. It is unusually direct. You do not need to create new machine-readable files, AI text files, markup or Markdown to appear in Google Search including its generative AI capabilities, because Google Search does not use them. A note added on 15 June 2026 addresses llms.txt by name: maintaining one is fine for other services that may consume it, and it will neither harm nor help your visibility or rankings in Google Search, which ignores it. There is no requirement to break content into tiny pieces for AI to understand it; Google's systems extract the relevant passage from a multi-topic page. You do not need to write in a specific way for generative AI search. And seeking inauthentic mentions across the web is not as helpful as it might seem. Four common line items, disclaimed at the source.

State the nuance, or make the opposite error

It would be easy to overstate this. The June note does not say llms.txt is harmful, and it does not speak for engines other than Google. Other services may read the file; it is cheap to maintain; the accurate description is that it is not a Google lever and should never be sold as one. The same care applies in the other direction. Google not using a signal is not evidence that nothing uses it. What the guidance establishes is narrower and more useful: any vendor whose pitch for the Google surface rests on AI-specific files, chunking or AI-specific rewriting is selling something the surface's operator has said it ignores. We removed our own llms.txt for that reason. Keeping a file we advise clients not to pay for would put us in the category this page describes.

What the controlled evidence rules out

The July 2026 survey of 45 studies collects the tactic-level results into one table. Keyword stuffing is null or negative across multiple benchmarks: avoid it. Formatting recipes on their own generalise poorly, with occasional local gains that do not survive a change of query or engine. An authoritative tone is weak and unstable as a lever and can conflict with credibility. Body rewrites are covered on their own page, because the largest end-to-end test found them negative. The levers with moderate support are the unglamorous ones: recent dates, explicit prices, specific facts on the commercial queries where a buyer would want them. If a proposal lists formatting, tone and density as its deliverables, the literature has already graded it, and the grade is not a pass.

The tool that would have had us lie

The most expensive item on a list like this is advice that breaks the law. On one property we measure, a well-known SEO tool's site audit flagged 25 structured-data errors on 14 August 2026. Twenty-one of them would have required adding star ratings and review counts the business does not have, because the tool's validator expects an aggregateRating on any page that carries product markup. In Australia, publishing a rating nobody gave is a false or misleading representation under section 29 of the Australian Consumer Law. The correct action was to fix the four real errors and leave twenty-one red marks in the dashboard. Any diagnostic that cannot tell a missing field from a fabricated one is not a diagnostic; it is a list of things that would make a dashboard go green.

Score your vendor

On 5 June 2026 Google added guidance on using third-party SEO tools, services and advice, including services sold as AEO or GEO. Its questions are the right ones. Is the advice aligned with Google's official guidance on generative AI features? Does the vendor cite official documentation as evidence for its recommendations? Does it use tools aligned with that guidance? The scorecard below combines those three questions with the tactic list above, each item carrying the source it came from. Answer it honestly for your current provider, print the summary, and take it to the next meeting. A good agency will welcome the questions and answer them with documents rather than adjectives. An agency that cannot is telling you something the proposal did not. That is, in the end, the test.