Why has your AI visibility stopped improving?

Visibility stops improving because the cheapest lever gets used up first, and the levers above it are slower, partly outside your control, and usually not in the scope you bought.

Page work is the cheap half. Rewriting a title into the question a buyer asks, answering it in the first sentence, making sections stand alone: those changes can be made in an afternoon and they move a readiness score immediately. What they cannot do is keep moving it. Once a page's language matches the question, there is nothing further to take from that page.

The expensive half is retrieval and familiarity. Chen and Koudas, studying 158,099 generative engine answers with 468,037 citations across two 90-day collections, attribute about 32% of the variation in citations to which pages the engine's own search retrieves and 2.8% to page properties (arXiv 2610.15033, 10 October 2026). They also report that ChatGPT writes brands it already knows into its searches, favours sites it knows, and names brands with no supporting page at all. Neither of those two channels is a property of a page you can edit.

Our own reporting makes the same split in plainer language. The early results page for our law firm work says the audit score is an input metric describing how well a site is structured for citation, while share of voice is the outcome metric, and that a strong audit score stacks the odds without guaranteeing a citation. A programme that moved the input and then flattened has done the first job and reached the second.

So the question to ask is not what went wrong. Ask which of three things the flat line is: noise, an exhausted lever, or work that was never started.

Is it a plateau or is it measurement noise?

Telling a plateau from noise takes repeated runs, and the published churn in AI citations is large enough that most reporting cadences cannot tell the difference.

Ahrefs analysed more than 43,000 keywords, each with at least sixteen recorded AI Overviews over a month, and found about 45.5% of cited sources were entirely new between consecutive captures of the same query, with the answer text changing about 70% of the time and an average cosine similarity of 0.95 between consecutive answers (published 11 November 2025, authored by Louise Linehan). Nearly half the citation list turns over between one observation and the next while the meaning of the answer stays put.

A second Ahrefs study puts a number on how long a citation lasts. Tracking about 37,500 citations by asking the same questions daily for twenty days across five tools, only 6.6% of day-one Google AI Overviews citations and 4.0% of day-one ChatGPT citations survived the full twenty days, between 50% and 83% of day-one citations were gone by day two, and roughly 75% of citations that disappeared at least once came back inside the window (published 29 October 2026). Gemini returned no citations at all in 64% of answers and Copilot in 53%, which inflates their apparent churn.

Two consequences for your report. A month-on-month comparison built from one reading per prompt is mostly sampling, and a dip followed by a recovery is the normal behaviour of the system rather than a result. Arithmetic sets the floor: a single share estimate at 95% confidence needs roughly 100 observations for an interval of plus or minus 10 points and roughly 400 for plus or minus 5.

Julius Schulte, Malte Bleeker and Philipp Kaufmann reach the same conclusion from the measurement side, writing that "Answers can vary across runs, prompts, and time, making one-off observations unreliable" and that visibility should be characterised as a distribution rather than a single-point outcome (arXiv 2604.07585, submitted 8 April 2026).

What runs out first when the work stops producing gains?

Alignment runs out first, and because alignment is the single strongest page-level lever, its exhaustion looks like the whole programme stalling.

Our own page is the clearest example we can show. We ran the SIGNALS audit on one of our own guides, which scored 28 out of 100, and rewrote it in an afternoon to 71, with nine content changes. Alignment went from 2 to 8 out of 10, almost entirely from rewriting the title and headings into buyer language. Substantiation moved from 1 to 4 and stopped there, because substantiation depends on third-party press and reviews rather than on anything that can be edited on the page.

That shape is not specific to us. Discovered Labs analysed about 2 million prompt, engine and citation observations over six months across ChatGPT, Claude, Google AI and Gemini, crawling 10,000 cited URLs with more than 60 features each, and measured prompt-content alignment at a standardised coefficient of +0.37, 95% confidence interval +0.33 to +0.41, significant at q below 1e-73 after false-discovery correction, which it describes as roughly 30% more citations for one standard deviation more alignment. The next strongest page-level feature in the same table is page length at +0.13, so alignment is about 2.8 times the size of its nearest page-level rival. Discovered Labs sells answer engine optimisation services, which belongs in any reading of its findings, and a predictor in that study had to survive four of nine robustness checks rather than all of them.

Read that as a stock rather than a flow. Alignment is a level a page either has or lacks, and a page whose title, headings and opening sentences already match how buyers ask has banked the whole effect. The second month of alignment work on the same page is worth a fraction of the first, and the fourth is worth nothing.

Why doesn't publishing another article move it?

Publishing another article adds coverage of a question you were not answering, which is useful, and does nothing for the questions you already lose. Those are different problems and they look identical on a visibility chart.

Mike King put the mechanism in one sentence: "Publishing another article before investigating often leaves the original issue untouched" (iPullRank, 4 November 2026). King sells agency services in this category, so read it as an argument from an interested party whose logic still holds.

There is a measured version of the same point about how much of your citation footprint your own pages can carry. Trakkr tracked a weekly panel of 1,050,171 AI answers from 2 March to 11 October 2026 and reported that pages a brand owns account for a median 4.6% of that brand's AI citations (study 011, first published 2 July 2026, updated 16 October 2026, by Mack Grenfell). Trakkr sells AI visibility software, and the figure comes from its first run rather than the rerun, both of which the study states.

Our own category research cuts the other way, which is why the honest answer is that it depends on the category. Across 27 buyer queries in three pharmaceutical skid lanes, 70% to 79% of the 4,395 citations behind the AI answers pointed at manufacturers' own capability pages rather than directories, trade media or third-party rankings. In a category with few independent publishers, your own pages are most of what there is to cite. In a category with mature review sites and trade press, they are a small minority.

So establish which kind of category you are in before you commission more content. Count where the citations in your own answers actually point. If they point at your competitors' own pages, more pages is a reasonable next step; if they point at review sites and publications, more pages will not move anything.

What is the next lever after the page work?

Four levers remain after the page work, and they are worth pulling in this order because each costs more and moves more slowly than the one before it.

Lever	What it fixes	How fast it moves
Single-engine failures	A zero on one engine while the others name you	Weeks, and the cheapest remaining fix
Unowned phrasings	Questions your buyers ask that nobody has answered well	Weeks to months, one page per question
Retrieval depth	Whether the engine's own sub-queries surface your domain at all	Months, through topical coverage and internal structure
Third-party presence	Being recommended rather than only cited	Months, and partly outside your control

Start at the top row, because a single-engine zero is the most fixable kind of invisibility and the easiest to miss. Our deep category study found one fabricator named in the top three on Google, Perplexity or Claude in twelve of the fourteen queries in its lane while scoring zero on ChatGPT in eleven of those same fourteen, which is one cause blocking an entire lane of queries on one engine.

The second row is about vocabulary rather than volume. SparkToro asked a panel of about 144 people to write their own prompts for the same goal and measured the average semantic similarity between those prompts at about 0.08 on a scale of 0 to 1 (published 8 May 2026, with research run by Gumshoe.ai, which sells tracking software). People asking for the same thing barely use the same words, so the gap between your coverage and your buyers' phrasing is usually wider than a keyword list suggests.

The bottom two rows are where a stalled programme normally has to go, and they are the rows most retainers were not scoped for. Our guide on third-party mentions covers how to tell whether that is your constraint.

How long should the next phase take before you judge it?

Give the next phase a quarter, decide the review date before it starts, and write down which number you expect to move.

Page work and single-engine fixes show up fastest, in weeks, because the published requirements are short. Google's documentation on its AI features states that a page must be "indexed and eligible to be shown in Google Search with a snippet", that "There are no additional technical requirements", and that "You don't need to create new machine readable files, AI text files, or markup to appear in these features" (read 9 November 2026). Nothing in that list takes a quarter.

Retrieval depth and third-party presence are the slow rows, and a quarter is the shortest honest window for either. Both accumulate: a page published this month contributes to the topical coverage that decides retrieval next quarter, and a mention earned this month sits in the pool a model reads from later.

Three rules make the judgement possible when the window closes. Measure the same prompt set you measured before, because a rewritten prompt set produces a new baseline rather than a result. Report every share with an interval, since a number without one cannot be compared with the last one. And separate the input from the outcome in the report itself, so a readiness score that improved while share of voice did not is visible as the ordinary thing it is rather than as a contradiction.

Our guide on how long AEO takes to work sets out the timelines in more detail, and the companion piece on telling whether AEO is working covers what a monthly report should contain.

When is a plateau a reason to stop paying?

A plateau is a reason to stop paying when three things are true at once, and a reason to change scope when only one or two of them are.

The first condition is that the page-level work is genuinely finished, not merely started. Finished means every page targeting a question you want is answering that question in its own language, your pages are reachable and readable by the crawlers, and the readiness work has nothing left to take. Most programmes that feel stalled at month three are not here yet.

The second is that the remaining levers are ones the provider cannot pull and will not say so. A retainer scoped for on-page work, billed monthly, against a constraint that now sits in public relations or product marketing, is a scope mismatch rather than a performance problem. The honest version of that conversation ends in a smaller retainer or a different supplier, and our guide on when to stop paying for AEO walks through it.

The third is that the number has been flat across enough observations to be a fact. Given roughly 45.5% source turnover between consecutive captures in the Ahrefs study above, a flat quarter measured once a month is four readings, which is not enough to establish anything.

Where all three hold, stopping is a reasonable decision and the work does not immediately unwind: the page changes stay, and the readiness they bought stays with them. Where the first two hold and you want to know which lever is left, a free visibility assessment runs your buying questions across ChatGPT, Perplexity, Google AI Overviews and Claude and shows which phrasings are open, which are owned, and whether your problem is the page, the retrieval or the brand.

What else do people ask when AI visibility stalls?

Marketers ask these five questions most often in the month after a visibility programme flattens out.