AI visibility tactics that do not work

What have we tried that did not obviously work?

Start with our own site, because it is the one we can show in full.

A long, thorough guide. The pillar page of our AI crawler guide is several thousand words and sourced line by line. In the 30 days to 14:45 UTC on 4 October 2026, when we re-read the page counts, no answer to the 61 questions we track attached it, though five of those questions are about AI crawlers (our measurement guides use the morning reading of the same day, 4,374 answers). Our pricing page, a fraction of the length, was attached dozens of times; both counts are at the top of this page. Length and thoroughness did not make the guide a source.

An llms.txt file and Markdown copies of every page. We publish both. Perplexity cited our llms-full.txt file in answers to "what is Attensira" in early October, so at least one engine reads it. We cannot show it changed what any engine said about us, and on every unbranded question we track we were named in no answers in that month.

Structured data on every page. Article, FAQ, breadcrumb and claim markup, generated from the same data as the visible text. It is correct and it costs nothing to keep. We have no evidence it earned a single citation.

None of these was wasted, exactly. They make our pages easier to read for any agent that comes looking, and they are the right defaults. But if we had stopped there and called it an AI visibility programme, we would have a well-built site that engines do not name. That is the trap this chapter is about: work that feels like progress and does not move the answers.

Why do so many AI visibility tactics fail?

Because they act on the wrong step. An engine names a brand when four things are true: it can fetch the pages, its retrieval finds them, a passage on them answers the question, and other sources it trusts agree. Most failed tactics push on a step that was not the constraint, or on no step at all.

The GEO study, the first systematic test of content changes against a generative engine, put it plainly: techniques that work in search engines may not translate. Its tests found some tactics helped and several did nothing.

Which content tactics do not work?

Keyword stuffing. In the GEO study, repeating target keywords gave little to no improvement in generative engine answers. An engine lifts a passage because it answers the question, not because it contains the question's words many times.

A more authoritative tone. The study also tried rewriting pages to sound more authoritative and persuasive. It found no significant improvement and concluded that engines are already somewhat robust to it. Confidence is not evidence.

Volume for its own sake. Mass-produced pages that answer nothing specific, such as hundreds of thin glossary entries or city-by-city landing pages with the same text, add pages engines can fetch and give them no passage worth quoting. They also give buyers who land on them a reason to leave.

Length as a strategy. Our crawler guide is the proof. A long page helps only if the passages in it answer the questions buyers ask, in sentences that stand alone. The passages chapter covers what does help.

Which technical tactics do not work?

Special files or markup as a ranking lever. Google says you do not need new machine-readable files, AI text files or markup to appear in its AI features, that there is no special schema.org structured data to add, and that there are no additional requirements or special optimisations. llms.txt is cheap and harmless; it is not a switch.

Separate pages for bots. Serving crawlers different content from people is an old risk with search engines and has no documented benefit with AI engines. One page, written for people, readable without JavaScript, serves both.

Blocking every crawler to "protect" content, then expecting citations. A single wildcard disallow removes you from every AI search product that honours robots.txt. And some user-triggered fetchers, Perplexity-User among them, are documented as generally ignoring robots.txt because a person asked for the page. robots.txt is a request to compliant crawlers, not a lock. Decide by agent and job instead; the crawler guide shows how.

Hiding instructions for AI in pages. Text aimed at manipulating a model, hidden from readers, is the content equivalent of cloaking. It risks being treated as spam by the engines that find it and as dishonest by the people who do.

Which off-site tactics backfire?

Fake reviews and undisclosed posts. Writing your own reviews, or posting in forums as an "unaffiliated user", can put your name on a page an engine reads. It can also get the review removed, the account banned and the thread quoted back at you. The pages engines trust most are the ones with rules against exactly this.

Getting added to lists where you do not belong. A place on an alternatives list for a product you do not compete with puts a misleading sentence about you on a page engines read. It may get you named in the wrong answer, to the wrong buyer.

Buying "guaranteed" AI placement. No assistant sells organic placement in its answers. Google says even a page meeting every requirement is not guaranteed to be crawled, indexed or shown. A vendor guaranteeing citations is selling something other than what the name says.

Which measurement habits mislead?

Watching rankings as a proxy for AI answers. A page can hold the top position while the AI answer above it names a competitor. The ranking chapter explains why.

One "visibility score". A blended score cannot tell you which ingredient moved. Attensira computes none and reports each rate with its answer count.

Counting crawler visits as visibility. On our site, Meta's training agent made about nine in ten AI crawler requests in a month and feeds none of the answer engines we track.

Single-answer screenshots. One answer is one draw from a system built to vary. A screenshot of a good answer proves as little as a screenshot of a bad one.

Revenue attribution from mentions. No tool can tie an AI mention to a sale; there is no identity link between the answer and the visitor. Claims that do are correlations with a causal label.

Which claims should make you suspicious?

Some sentences in this market are reliable signs that a tactic, or a vendor, has not been tested honestly. Ask for the evidence behind any of these:

• "Guaranteed citations" or "guaranteed rankings in ChatGPT". Nobody controls what an engine says. A guarantee can only be about work done, never about the answer.
• "Our AI visibility score went up 40 points." Ask what the score is made of, which questions it covers and how many answers sit behind it.
• "One simple file gets you into AI answers." No operator documents a file that does this. Google says its AI features need none.
• "We saw it in ChatGPT." One answer, once, is anecdote. Ask how many answers, on how many days.
• "AI drove this revenue." Ask how the answer was joined to the sale. There is no identity link to join it with.

Attensira's own promise is narrower on purpose: it commits to the work, not the outcome, and reports every result with the answer count it came from. Hold any vendor, including us, to that standard.

How do you tell whether a new tactic works?

Test it the way you would test anything else, on a small scale first.

1. Name the step it acts on. Fetch, retrieval, passage or corroboration. If you cannot name one, it is probably decoration.
2. Pick the questions it should move, before you start, and write them down.
3. Measure before, for long enough to know the normal range: at least a few weeks of daily answers on each engine.
4. Ship it on some pages, not all, so you have a comparison.
5. Measure after, at one, two and four weeks, and only call a change real past a significance test.
6. Report the null. "Nothing we can prove yet" is a result. Most tactics, honestly tested, produce it.

That last point is uncomfortable and important. In a field full of confident claims, the most useful thing a team can do is keep a list of what it tried and what the answers did afterwards, including when they did nothing.

What does work, then?

The short version of this guide:

• Let the search and user agents in, and serve the answer in the HTML.
• Write passages that stand alone, name your brand, answer first and carry a specific.
• Earn an accurate place on the third-party pages that already decide your buyers' questions.
• Measure per question and per engine, with answer counts, and judge changes over weeks.

None of it is a trick. It is the same work an SEO agency would do, aimed at a new kind of reader. That is the work Attensira's crew does every day. Workspaces start in Draft, so nothing ships until you approve it unless you choose otherwise, anything that states a claim always waits for you, and every shipped change is re-read so you can see what it moved.