What does not work

Which tactics are a waste of time, and how do I tell?

Three tests, applied in order, dispose of most of the advice in this category.

1. Is there an operator statement? If a platform documents the behaviour, that

is the strongest evidence available.

1. Is there a measured result? One controlled study exists at scale. Its null

results are as useful as its positive ones and are quoted less often, for obvious commercial reasons.

1. If neither, is the claim at least falsifiable? "Improves your AI presence"

is not. "Increases citation rate on this prompt set from 4 of 20 to 9 of 20" is.

Almost everything below fails at step one or two, and the reason it survives is that it fails step three too - it cannot be checked, so it cannot be disproven.

Measured not to work

These are null results from the GEO study, not opinions.

Keyword stuffing. Adding more query keywords to page content "offers little to no improvement on generative engine's responses."[^geo-keyword-stuffing] The study's own conclusion is that owners need to rethink their strategies for generative engines because "techniques effective in search engines may not translate."[^geo-rethink] This is the single most transferable habit from SEO and it is the one that does not transfer.

Authoritative tone. Rewriting content to sound more persuasive and authoritative produced "no significant improvement", and the authors conclude that generative engines are "already somewhat robust to such changes."[^geo-authoritative-null] Their recommended redirection is explicit - focus on content presentation and credibility instead.[^geo-focus-shift] Marketing voice is not a retrieval signal.

The pairing is instructive. Keyword density and confident copy are the two levers a traditional content team reaches for first, and neither one moves the measured outcome. What moved it was adding a statistic, a quotation, or a citation.

Documented not to be required

Not the same as "harmful". These are things sold as prerequisites that an operator has said are not.

llms.txt. Google: "You don't need to create new machine readable files, AI text files, or markup to appear in these features."[^google-no-ai-files] No major operator currently documents the file as a retrieval input. It costs almost nothing to publish and it might matter later; what is not defensible is selling it as a mechanism, ranking it above technical access in a plan, or reporting its existence as progress.

AI-specific structured data. The same sentence rules out special schema.org markup for AI features. Structured data remains worth having for the reasons in chapter five; "for AI" is not one of them, on the one platform that has said anything.

A separate AI optimisation programme, on Google. Google states there are "no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary."[^google-no-additional-optimisation] Read that as the scope statement it is: on Google, AI visibility work is site quality work. The genuinely distinct work is on the platforms Google does not run.

Unevidenced in either direction

The honest category, and the one most guides quietly convert into assertions.

• Whether AI-written content is penalised in citation. No operator publishes a

detector-based demotion for AI answers. Anyone claiming a percentage here is reporting a correlation from a sample they chose.

• Whether a specific schema type raises citation rates. No published mechanism.
• How quickly a change propagates to a specific assistant. Only Google says

anything, and only about recrawling.

• Whether posting on any particular social platform influences citations.

Plausible via indexing, undocumented as a mechanism, and heavily promoted by people who sell posting.

The correct treatment of this list is to test it on your own prompt set if you care about the answer, and to state the absence of evidence when you report on it. That is what Attensira's own reporting rule is for - distinguishing "we did not measure this" from "we measured zero"[^attensira-unmeasured] - and it applies to advice as much as to metrics.

Actively harmful

Invented statistics. A page cited for a number that does not exist in its source is worse than a page nobody cites, because the error propagates into answers and outlives the correction. This is the failure this guide is built to avoid, and avoiding it is structural rather than aspirational: in this repository, a percentage that appears in a page's prose with no matching sourced entry fails the build.

That gate exists because we needed it. An earlier generation of pages on this site carried hundreds of fabricated statistics - confident, plausible, unsourced, several attributed to research nobody had run. They were purged, and the schema was changed so that a claim without an absolute source URL and a retrieval date cannot merge. If a vendor's guide has no such story, either they have not looked, or their content is younger than the problem.

Fabricated ratings and reviews. Emitting AggregateRating markup with no real ratings behind it is fabrication in a format designed to be trusted. It is also a structured-data guidelines violation, so the downside is not hypothetical.

Blanket bot blocking as a reflex. A User-agent: * disallow removes you from answer surfaces along with training crawlers, and does not even reliably stop user-initiated fetching - Perplexity states that its user fetcher generally ignores robots.txt because a person requested the page.[^perplexity-user-ignores-robots] Blocking is a legitimate choice made deliberately, per agent. As a reflex it costs you the thing you were trying to protect the value of.

Publishing volume without answers. Pages that exist to have been published compete with each other, dilute the site's topical focus, and answer no question a buyer asked. The measurable levers in the published research are all per-page substance.

Tactics that are not wrong, just mis-ranked

A third category deserves separating out, because calling these "myths" would be unfair and treating them as strategy is how quarters disappear. Each one is real work with a real, small effect, routinely placed above work with a large one.

Rewriting metadata. Titles and descriptions matter for click-through in classic search. They are a small part of what an answer is assembled from. A sitewide metadata sweep is a week that could have gone into the source list from chapter four.

Adding an FAQ block to every page. The shape helps where the questions are real. Bolted onto a page that has no questions, it produces invented queries and padding - which is the slop signature, not the citation signature.

Chasing every new platform. Each additional platform you sample divides your measurement budget. Two platforms measured deeply beat five measured shallowly, because only the deep ones can detect a change.

Publishing a glossary. Useful, cheap, genuinely extractable - and it answers definitional questions, which are the least commercial questions in the set. Worth having, never worth being the plan.

Refreshing dates without refreshing content. Bumping a date with no real change is worse than a stale date: it trains anything reading your timestamps to ignore them, and spends the one signal you had for a genuine update.

The common thread is that all five are things a content team can complete alone, on schedule, with a visible artefact at the end. The high-value work - access, corroboration - requires someone else's cooperation and produces no artefact. That asymmetry, not ignorance, is why the low-value work keeps winning.

What we got wrong

Two admissions, because a chapter like this is worthless from an author with a clean record.

We ran leaderboard-style pages at scale - a brand-visibility engine crossing thousands of brands with dozens of countries. It worked as SEO and failed as marketing: the traffic was people looking up a ranking, not people evaluating a tool, and the generation cost was thousands of dollars a month. It is switched off. The lesson we encoded afterwards is that a page which cannot plausibly convert is a cost, not an asset, however well it ranks.

And we published those fabricated statistics. Not maliciously - they were generated in bulk, they read plausibly, and nobody checked. The response was to make checking structural rather than cultural, because cultural checks fail exactly when volume is highest.

How to evaluate the next tactic you are sold

This field generates new tactics faster than anyone can test them, so the useful skill is triage rather than a list. Four questions, in order, and a tactic that fails the first two is not urgent no matter how confidently it is presented.

Who says so, and where can I read it? An operator's documentation page beats a vendor's blog post, which beats a conference slide, which beats a screenshot on social media. If the chain of sourcing terminates in someone's assertion, you have found the bottom.

What would falsify it? A tactic that predicts nothing specific cannot be wrong, which means it also cannot be right. "Improves AI visibility" predicts nothing. "Raises citation rate on comparison questions" predicts something you can check against your own prompt set.

What does it cost, including attention? Publishing an llms.txt costs ten minutes and no ongoing attention, so its weak evidence base barely matters. A programme of monthly schema audits costs a person, and its weak evidence base matters enormously. Judge the two differently.

What is it displacing? This is the question nobody asks and the one that decides outcomes. Every tactic is competing against reading your crawler logs and fixing your directory entries. Most lose that comparison badly, and they win anyway because they are easier to schedule.

Applied honestly, these four dispose of most of what appears in this category each quarter - including, on a bad quarter, some of what we would like to sell you.

What to take from this chapter

Ask for the source. Every recommendation in this field is either an operator statement, a measured result, or someone's guess, and the three are easy to tell apart once you insist on the distinction. Note that even the strongest evidence is bounded: Google's own documentation ends with the reminder that meeting every requirement does not mean it will crawl, index or serve your content.[^google-no-guarantee] Anyone promising more certainty than the platform itself offers is selling something.