Citation tools are fundamentally different from rank trackers, and that difference is almost always seen or stated as a limitation.
One respondent to my recent survey of digital marketing practitioners put the case plainly. You cannot reverse-engineer what is working when the answer changes every time you ask, so what you are left with is closer to a brand awareness signal than a diagnostic. I have heard some version of that from enough people now that it functions as the default reading in this space, and it is a fair one. That survey deliberately presented what respondents said without rebuttal, so I cut my response at the time. This is the response.
Their POV got me thinking, and what follows is where that thinking has led me so far. I will say again, up front, that I run CitationIQ, an AI optimization data platform, so I have a commercial interest in the answers to things like this. Feel free to discount me accordingly.
The Question I Am Asking Is Not The Same One
The default reading assumes the job of the tool is to explain why you did or did not appear. That expectation comes straight from rank tracking, where the position was the outcome, and the diagnostic work was figuring out what moved it. Wanting that back is reasonable. A number you can act on is more useful than a number you can only observe.
I have come at it from a different question. Not why you appeared, but whether the phrase you were chasing is still contested or has settled. Those are two different things to want from the same evidence, and I have landed on the second one being the one that matters commercially now.
Because if the answer has settled, and the answer is not yours, the diagnostic question has already been answered in a way no amount of reverse-engineering will improve on. The end user does not care whose answer it is. They wanted the answer, they got it, and the identity of the source was never the point for them.
Settled Is A More Useful Word Than Ranked
Here is what I think is happening.
Traditional SEO treated phrasing as expandable. There were many ways to ask the same thing, each one countable, each one a separate opportunity, and the whole method was aggregating those variations into volume worth chasing. The new systems treat that same phrasing as collapsible. They take the variations, average across sources, and return one answer that the person accepts and acts on. That is what I mean by convergence.
If I have that right, it is close to an inversion. The thing the industry spent twenty years expanding is the thing these systems are built to compress.
But convergence is not as clean as that makes it sound. Models do not reliably land on one answer. A June 2026 audit of 3,750 responses across three models and 250 category queries found all three agreeing on the top brand only 41.6% of the time. The more useful number from the same audit is the one underneath it. Majority agreement, where at least two of the three named the same top brand, reached 91.6%.
So the models are settling on which brands are eligible, not on which one comes first. The set is small and stable. The order moves around. When someone reruns a prompt and gets a different top answer, that is movement inside a fixed set, and treating it as proof that nothing has settled reads the wrong layer.
That changes what a citation tool is telling you. Not your position, which was never stable and never will be, but whether the phrase still has room in it. If 10 queries you treated as 10 opportunities all resolve to the same short list, they were one opportunity, and now you know.
Someone will say this is the featured snippet debate again. It is not, though the economics are similar. Snippets collapsed the click and left the answer space intact. The phrase stayed contested, one publisher held the box, and you could see who held it and go take it. Ahrefs measured the damage at the time. Convergence works differently because the answer is built from several sources at once, so there is often nobody holding anything to take. Practitioners who say they have seen this happening before are right about the effect and wrong about the mechanism, and the mechanism is what decides whether the old response still works.
But Is Any Of This Real?
The strongest objection is that convergence is an artifact of how it gets measured. Clean sessions, synthetic prompts, no user history. If every real user gets a personalized experience, convergence might be something that only exists inside a test space.
Personalization does not appear to dissolve convergence. It appears to relocate it. An audit of 2,000 runs across ten buyer personas found category leaders largely persona-resistant, holding roughly 80% consistency regardless of who the model thought was asking, while mid-market brands swapped up to 75% of the recommendation set as the persona changed. The leaders stay put no matter who is asking, and the churn happens below them. Which means personalization concentrates the problem I am describing rather than solving it.
The synthetic prompt objection I cannot answer as cleanly. Nobody in this category, including me, is currently measuring against verified real-world query distributions at scale. That is a real limit on what any tool here can claim right now, mine included. (And scale here refers to “all of it” not “we sampled 1,000,000 instances and found X”. Nice, but only a fraction of the overall.)
The Map Has Fewer Places On It Than We Assumed
Google documents that AI Overviews and AI Mode may issue multiple related searches across subtopics and data sources before building a response. So the phrase a person types is frequently not even the phrase the system searches. That is the compression happening one layer earlier than most people are looking for it.
Here is the part that will be unpopular. The space of genuinely distinct commercial opportunities was always smaller than the space of phrasings. Convergence did not shrink it. Convergence made it visible.
I watched a version of this from the inside. During my years at Bing, category-level attention concentration was well understood, and it shaped where resources went. Entertainment, autos and news drew people and server capacity because that is where the aggregate demand sat. Categories like sewing or knitting mattered enormously to the people they mattered to, and got proportionally less. That is ordinary resource management applied to information retrieval, and it was true two decades before anyone trained a language model on the open web.
What is new is that the concentration now decides answers instead of just budgets. Researchers at Trine University and Texas A&M ran an experiment. They built product sets of one real brand against nine validated fictional ones, with identical ratings, prices, review counts, and ingredient descriptions. The only difference was the name. The real brand was recommended in every one of 670 valid trials, across three models, two languages and four product categories. Not once did a fictional brand surface.
The model was not evaluating products. It was recognizing a name. Which tells you what winning looks like now, and it is not being the best answer. It is being the most described entity in a category where description has already accumulated. The same June audit found genuine competitive vacuums, meaning category queries with no dominant brand at all, in only 8% of 250 queries.
I have gone in-depth on trust in earlier articles. The point worth pulling forward is that these systems need reliable sources, because a synthesized answer is only as good as what it was built from. Recognition is the cheapest proxy for reliability available, so the models lean on it. None of that should shock anyone. What is shocking is knowing all this and still deciding that not being able to see rank is the problem that needs solving.
Why The Industry Would Rather Not Look At This
Fewer distinct opportunities means fewer businesses can win, and the ones that do will win on something other than phrase coverage.
That is an existential reframe for a discipline whose economics assumed everyone could eventually find their niche. The long tail was never only a tactic. It was the promise that there was room for everybody, that a small operator with patience and a content budget could build something defensible. Facing convergence honestly means facing a smaller addressable opportunity than the one a lot of careers were built on, mine included.
I do not think practitioners are avoiding this out of bad faith. The incentive not to look is completely understandable, and I did not arrive at it cheerfully myself.
One more thing complicates the picture. Nearly all the published measurement of AI brand visibility comes from companies selling AI brand visibility measurement. Two of the three studies above are vendor research with disclosed conflicts. That is the same conflict I declared about myself, showing up across the entire evidence base, and it is a reason to hold every number in this piece loosely.
Where This Argument Runs Out
Convergence may be temporary. Retrieval architectures change, model families diverge, and today’s canonical consideration set may fragment again in 18 months. I have no way to predict that risk.
The bigger limit is query type. Everything noted above is strongest for informational and category-level questions and weakest for specific commercial ones. Convergence on what is X tells you very little about best X for Y under constraint Z. The cross-model agreement data cuts against me as much as for me here, because the 41.6% figure came from commercial category queries, which is precisely where my argument is doing the most work and carrying the least support. If this only holds for informational queries, it matters considerably less than I think it does. I do not believe that, but I cannot rule it out on what has been published so far.
So What Replaces Phrase Coverage?
I do not have this fully worked out yet, and I’m hoping to hear your thoughts on it.
I think some directions look more promising than others. Being the source models converge on, rather than one more source competing for a phrase, is the obvious one and also the hardest, because it is earned through independent description over years rather than produced on a content calendar.
Entity-level standing rather than page-level optimization follows directly from the recognition finding. If the model is selecting on the name it knows, then the unit of investment is the name, not the page.
Categories where convergence has not happened yet are real, and the mechanism tells you where to look. Kandpal and colleagues established that a model’s ability to answer about something tracks how many relevant documents it saw during pretraining. Mallen and colleagues found that scaling improves recall at the popular end while leaving the sparse end roughly where it started. Sparse categories are where the vacuums sit, and healthcare technology showed the highest vacuum rate in that June audit at 20%. Thin coverage is an opening, and a temporary one.
And some queries are simply not winnable and should be abandoned rather than fought. That is the least satisfying item on the list and probably the most valuable, because the cost of contesting a settled phrase is not only the wasted spend. It is the phrase you did not contest instead.
What I keep returning to is that convergence is not a failure of measurement. It is a measurement of something this industry has not had a way to see before, and what it appears to be measuring is how much room is left, I think. That is uncomfortable. A smaller map you can actually see still beats a large one you were imagining, however.
If you are testing this against your own data and getting a different answer, I want to hear about it. Leave a comment below or reach out directly.
I go deeper on how these systems build and hold their picture of a brand in The Machine Layer, available here.
More Resources:
This post was originally published on Duane Forrester Decodes.
Featured Image: dotshock/Shutterstock; Paulo Bobita/Search Engine Journal
