Why your agency's AI visibility work is not producing results

Most of what agencies ship under the heading of AI visibility operates on pages the client owns, and answer engines mostly quote pages the client does not own. In Salienz's AI Security Shortlist Study of 7,385 answers from seven AI assistants, vendors' own websites took 22 percent of the citations in the answers that recommended vendors. Roundups and comparison pages took 40 percent, and review sites, forums and publishers most of the rest. 3 Blog posts, schema markup, an llms.txt file and on-page copy edits are all real work. None of them changes what a third-party roundup, a review platform or a community thread says about you.

Some of that on-site work does not move recommendations even on its own terms. Salienz's crawl of 99 cybersecurity vendors' websites found that vendors with an llms.txt file were named in a median of 0.8 percent of answers, the same as vendors without one, and FAQ markup showed no link either. 4 A larger B2B SaaS study of roughly 350,000 articles found the same pattern from the other side: earned media carried 61 percent of AI citations and brand-owned content 29 percent. 5

That is the first thing to check, and it is checkable without an argument. Ask for the last quarter's deliverable list and sort every line into two piles: things published on a domain you control, and things published on a domain you do not. If the second pile is thin or empty, the work is aimed at the wrong surface, and no amount of doing more of it will move a citation rate.

An SEO agency improves pages you own so a search engine ranks them. An AI visibility agency works the sources an answer engine quotes, most of which are pages somebody else owns. The disciplines share a technical foundation and diverge completely after it.

Read the right-hand column and the diagnosis usually writes itself. An agency whose entire quarter sits in the top three rows is doing SEO, competently or otherwise, and calling it AI visibility. That is a scope problem rather than a competence problem, and the two get confused constantly because the work itself is often good.

The technical rows are not worthless. If an engine cannot crawl or render your pages, nothing downstream matters, which is why crawlability sits in the table as a prerequisite. What it is not is a thing you keep buying. It is fixed once and then verified.

The difference is the deliverable, not the label. SEO services are sold in articles, technical fixes and links to your own domain. AI search optimization services are sold in placements, mentions and profile work on domains you do not control, plus the measurement to prove any of it happened.

The two are frequently sold under one another's names, in both directions. An SEO retainer relabelled as AI visibility is the common case. The rarer and more expensive case is an AI visibility retainer that quietly stops doing the off-site work when it gets hard, and backfills the report with content volume, because content is the deliverable that is always available to produce.

One question separates them on a call: how many placements on domains we do not own are committed per month, and what happens to the fee if they do not land. A vendor selling off-site work answers with a number. A vendor selling content answers with a process. Neither answer is dishonest; they describe different purchases. What the published rates for each look like is in what GEO agencies actually charge per month. 2

You need the function, not necessarily a second vendor. If your current agency will commit to a monthly placement count on third-party domains and report citation share against named competitors, they can do this. If they will not commit to either, adding budget with the same vendor buys more of what already is not working.

• Keep one agency when they already run digital PR or outreach, and the change you need is where the work points rather than what the team can do.
• Keep one agency when your category is narrow enough that the target list is short, because relationship depth beats headcount on a short list.
• Split when off-site placement would be a new capability for them, since a team learning outreach on your budget is slower than one that already has the relationships.
• Split when the reporting conflict is unresolvable, because an agency measured on sessions has no reason to prioritise a mention that produces a citation and no click.

The reporting conflict is the one people underrate. An agency paid against traffic will rationally deprioritise work that produces a citation nobody clicks, and that citation is the entire point. The opposite mistake is just as common: a report that counts citations of your website and calls it visibility. In the same study, one vulnerability management company's site was cited in 199 answers across seven assistants, mostly its guides, while the company was named in 9. 3 Ask for mentions of your name and citations of your site as two separate numbers. Changing the metric sometimes fixes the agency without changing the agency. What a placement-committed scope looks like written down is on our services page, which is at least a template for what to ask yours for. 1

Before firing an agency, ask for four things in writing. The answers take a day to produce if the work is real, and the absence of any one of them is more informative than any argument about performance.

1. 01The deliverable list for the last quarter, split by whether each item published on a domain you own or one you do not.
2. 02The baseline: which prompts were tracked, on which engines, how many runs each, and what the naming rate was on day one. Fewer than about 90 prompts cannot show a change of a couple of points.
3. 03The current naming rate on that same prompt set, measured the same way, so the two numbers are comparable.
4. 04Which named competitors gained or lost citation share over the same window, because if everybody moved together the cause is the engine and not the agency.

An agency that cannot produce the second item never established a baseline, which means nothing they have done since is attributable in either direction. That is the single most common finding, and it is not necessarily evidence of bad work. It is evidence that neither of you can tell. Size matters as much as existence: in Salienz's Audit Size Test, two readings of an unchanged brand built on 15 questions differed by 2 points or more in up to 44 percent of draws, against 7 percent or less with 90 questions. 6 A small baseline and a small re-check can show a gain, or a loss, that is only noise. How big a check needs to be is set out in when you need an AI visibility audit, and when you need a retainer, and what one costs in what an AI visibility audit costs.

If you do not have a baseline and want one from somebody with no stake in the answer, that is what an $800 audit is for. Its most useful possible outcome is telling you your current agency is doing fine.

Weeks for a page to be picked up, if it is picked up at all, and a quarter for a result you can read. In Semrush's test of 81 new pages on its own blog, ChatGPT search cited 8 the day after publication, 14 by day 7, 28 by day 14 and 34 by day 30. Google AI Mode cited 29 within a day, peaked, then dropped pages again and was citing 21 at day 30. 7

Those were pages on one of the strongest domains in marketing, written for the purpose. A placement on somebody else's site still has to be published, crawled, indexed and retrieved for a relevant prompt before it is quoted, and every one of those steps runs on somebody else's schedule. Two lessons follow. Most new pages do not get cited, so a quarter of placements is a quarter of attempts rather than a quarter of guaranteed sources. And citations come and go from one day to the next, which is why a quarter, measured on about 90 questions, is the shortest window over which a change can be told from noise.

It is also why the leading indicators are worth watching separately. Watch pages indexed, then pages retrieved into an engine's working set, then pages actually cited. Movement in the first two with nothing in the third is a real signal, and it means extractability rather than reach is the problem.

What should worry you at week six is not the absence of citations. It is the absence of placements, because those are entirely within the agency's control and are the input everything else depends on.