Here's the version of this story that gets told most often: someone gets frustrated with the tools available, spends some evenings building their own, and ends up with something neat to show people. Charming. Also incomplete, because it skips the actual decision underneath it — the one that happens before a single line of code gets written.
The gap has to be real first
To be fair to most of the big tools & the established brand names on the market they have built AI-visibility features to keep pace with the changing nature (or should that be changing appearnce?) of search — this isn't marketing dressed up as development. The problem sits a level deeper than that. They built those features using the same underlying measurement philosophy the industry has always used: are you appearing, counted, present. That's position-counting with a different name attached to it — the same proxy-metric instinct that made rankings the currency of every client conversation for twenty years, just pointed at a new target. Whether a client's actually better off for the citation is a separate question the tool was never built to ask.
The tool you use dictates the conversation you have. If citation count or "AI visibility score" becomes the headline number, that's the conversation — same as when rankings were the headline number. And it skips the two questions that should come first: does this client actually want to be visible for this, and why — what are they expecting to get from it. Being cited today doesn't mean anything tomorrow if it was never connected to an outcome that matters to the business. A metric that's faulty or non-representative of the bottom line doesn't become useful just because it's now measuring AI answers instead of blue links.
That's the actual test for whether a gap is real: not "does a version of this feature exist somewhere," but "does the tool measure something that predicts the outcome, or just something that's easy to count." If that gap isn't real, building is just an expensive way of reinventing something that already works fine. The strategic judgement is in telling the difference.
The build-vs-buy question, one level up
It's worth remembering this isn't just a question individual consultants and agencies face. The AI labs are living the same decision at a scale that makes everyone else's version look small. Very few companies actually have a web-scale index of their own — the crawl infrastructure, the storage, the years of indexing history it takes to build one properly. Most AI answer engines are, in some form, renting that layer rather than owning it: licensing search access, building on top of an index somebody else spent a decade constructing, because building their own from nothing was never realistic on any reasonable timeline.
Which means Google isn't just shaping SEO reporting when it tightens access — covered in the previous piece on this site, the same week Google quietly gated full-web search results behind a partner-only, unpriced API. It's shaping how much genuine independence any AI answer engine can have from the infrastructure Google controls. The companies positioned to build real competing answer engines are the ones who made the build call years ago, when it was hardest to justify. Everyone else is buying access to somebody else's decision, on terms that can change without warning — which is exactly the position the free side of that gate is in right now.
The conversation you can't stop having
Rankings aren't going anywhere in the client conversation, whatever the tooling does. A client understands "you're position four for this term." They don't yet have an intuitive feel for citation share, and most agencies are still translating AI visibility back into ranking language just to be understood — which quietly keeps rankings as the actual currency of the relationship, no matter what the reporting dashboard says underneath. (Worth noting this isn't speculation — John Mueller, Google's own Search Relations team lead, confirmed directly this month that there's no fixed, stable way to count position for AI Mode and AI Overviews yet, that it's expected to keep evolving as search itself does, and that comparing these metrics now takes real interpretive work on the site owner's end. That's Google itself saying the ground is still moving.)
There's a tell worth noticing here. If an agency says they're "constantly having to explain GEO to clients," that's not evidence the client is slow. It's evidence the explanation hasn't landed yet — because a genuinely settled answer stops needing to be re-explained every time. Spend the whole conversation re-litigating what GEO even is, and there's no time left to actually optimise for it. The explaining becomes the work, instead of a five-minute preamble to it.
That's not a client-education problem. It's a sign the answer being given isn't actually settled in the agency's own head yet — because if it were, it would fit in a sentence, not a recurring meeting. Realistically, this argument should have been settled years ago, when mobile search first splintered a single 'ranking' into countless different results — by device, by location. Then stack the personal layer on top of that: Google's had the best part of twenty years of accumulated personal search behaviour to draw on, and AI platforms are building the same kind of profile now, just faster. Incognito mode was never really incognito, not in any way that matters here. Wisdom of the crowds only works if the crowd is actually seeing roughly the same thing. Once results are this individually filtered, there's no reliable 'the crowd' left to be wise about — a glass with no fish in it doesn't mean the ocean's empty, it means you're looking at your own glass, not everyone else's.
Worth pointing at the sidebar here rather than just asserting it. Reddit held roughly 60% of ChatGPT's citation share, then dropped to around 10% in six weeks — not because Reddit's content got worse, but because Google removed the num=100 parameter in September 2025, the same change covered in the previous piece on this site, and it rippled straight through to which sources an entirely different company's product was citing. If a client's visibility can swing that hard from an infrastructure change nobody at either company chose to announce loudly, citation count was never a stable enough number to build a strategy on its own.
Worth being direct about the wrong lesson here too, because it's the obvious one to reach for. Reddit dominating citations right now isn't an instruction to go and spam Reddit — that's the same mistake as chasing whatever loophole Google happened to reward at any given point over the last twenty years, instead of understanding why it worked and what replaced it once the platform matured. If you want a read on where ChatGPT citations are heading, look at what Google itself did roughly eight years ago: crack down on manipulation, push harder on genuine expertise and quality signals, penalise anything that looked like gaming the system rather than earning it. AI platforms are almost certainly walking the same road, just faster. The takeaway isn't "exploit today's gap." It's "today's gap is temporary, and the platform is going to mature past it — same as it always has."
Domain Authority belongs in the same conversation, and it's worth naming directly rather than working around it. It's not even a Google metric — Moz built it as their own scoring model — so "well, just Google it" doesn't actually settle anything when a client pushes back with it, and correcting that rarely lands as useful in the room either way. It's one more number that's easy to cite, easy to screenshot, and doesn't predict anything real about the business. Using it doesn't make the client conversation easier. It makes it harder, because now there's a second proxy metric to walk back once the real numbers don't move the way the score implied they should.
Why building beats buying, here specifically
Off-the-shelf AI-visibility tools mostly do one thing well: they run a fixed set of prompts on a schedule and tell you whether you got cited. Useful, as far as it goes. What it doesn't do is tell you why — whether a site's structure is even legible to a crawler that doesn't execute JavaScript, whether the entity relationships across a page are clear enough for a machine to extract meaning from, whether the underlying architecture is the actual reason nothing's being cited in the first place. Buying a tool gets you the symptom. Building one gets you the diagnosis, because you built it to answer the question you actually had, not the question a vendor decided was sellable to everyone at once.
That's the trade being made, stated plainly: a bought tool is generalised, fast to get, and stops exactly where its designer decided the average customer's curiosity would stop. A built one costs more time upfront and keeps going exactly as far as the actual problem needs it to.
What it unlocks, once it's real
A tool built to answer your own question properly tends to answer several adjacent ones you didn't set out to ask. Mapping entity relationships to understand AI-citation gaps also surfaces every structural weakness underneath — the same audit work that improves traditional crawlability, feeds better internal linking, and gives conversion and UX work a genuinely evidenced starting point instead of a guess. None of that was the brief. All of it showed up anyway, because solving the real problem properly tends to solve several smaller ones standing next to it.
Worth putting this in the terms a client actually cares about, which is money charged or money saved, not features. It comes down to scale and value, and they're doing two different jobs. Scale is speed and reach — answering a question properly, at size, faster than piecing the same answer together by hand across two or three separate tools that were never built to talk to each other. Value is what makes the fast answer worth anything: the experience built into how it's interpreted, not just generated. A fast wrong answer isn't worth more than a slow right one. Together, that's what actually moves a campaign faster — not the tool itself, but scale and value pulling in the same direction.
The honest limit is resourcing, not thinking. A tool built and run by one person, on their own infrastructure and their own budget, hits a ceiling that has nothing to do with whether the idea was right — compute, storage, the paid-tier data that a wider rollout needs. That's not a flaw in the strategy. It's just the part of the strategy that needs more than one person's resource to finish properly.
The counter-argument, and where it actually lands
Worth engaging with the other side of this properly, not just citing agreement. Kevin Indig and Amanda Johnson published a piece the same week this one went up, citing MIT research on enterprise AI projects: bought and partnership tools reach deployment roughly twice as often as in-house builds, and the majority of internal build attempts fail outright. Their advice is direct — buy the tool when the problem isn't unique to you, and they name rank tracking, citation monitoring, and crawl diagnostics specifically as exactly that kind of common problem.
That's not a soft challenge, it's a sharper bar than "does the gap feel real." Their actual test is whether the workflow is genuinely yours alone, and — the part worth taking seriously, without qualification — whether someone on the team can verify the tool's output by hand. An AI tool nobody can check hands back wrong answers with total confidence, and that risk doesn't go away just because you built it yourself instead of buying it. No disagreement there. A pair of human eyes on the output is never optional, whichever side of build-vs-buy the tool came from.
Worth being precise about their own example, though. Amanda's story — a model misreading a GSC query spike as proof of rising AI visibility, nearly reaching the head of marketing before anyone caught it — gets told as a story about the model being wrong. It's really a story about the process being wrong: there was no review step until after the near-miss. The model didn't fail on its own. The workflow around it had no verification gate built in yet. That's the same point, made sharper — the fix isn't a better model, it's building the check in from the start rather than adding it after something nearly goes out the door; it's only a mistake if it leaves the kitchen.
The disagreement is narrower than "citation tracking is common, so buy it." Most crawlers already flag noindex tags, robots.txt blocks, and JavaScript-rendering gaps — none of that is new, and none of it needed building from scratch on its own. What doesn't exist off the shelf is a tool that ties those same findings together into one connected story: this noindex tag isn't just a crawl error, it's why the page's entities never reach the content sitting underneath it, which is why nothing built on that page gets cited. Bought tools report each of those as a separate, piecemeal flag on a list. Built for the actual question, they're one causal chain, and a much easier one to explain to a client in a sentence rather than three unrelated line items.
That's the honest version of the claim, and it's a narrower one than "build beats buy": not that the problem is unique, but that the connected version of the answer isn't available anywhere for sale yet.
Worth addressing the implied worry too — that a self-built tool only the builder understands is a knowledge-transfer risk a bought one avoids. It isn't, and SEMrush (as an example, there are countless others) proves the point rather than undermining it: nobody opens it and understands what the numbers mean without training either. Every tool needs an instruction manual, bought or built. The difference is what's written into that manual by design. Built for this specific question, the logic itself gets coded in — heavy JavaScript is a problem because it blocks AI visibility, not just a performance note; a non-indexable page in the main navigation isn't a minor crawl flag, it's the reason nothing downstream of it will ever be seen. That connection gets made automatically, at the point of the audit, instead of being discovered months later after the spend that could have avoided it. A raw flag nobody knows how to interpret might as well not have fired at all.
Buying a tool doesn't remove this risk either, it just relocates it. You're trusting someone else's interpretive logic instead of your own, and assuming their pipeline never breaks — which is its own leap of faith. Even Google's own first-party data isn't exempt: Search Console has had its own well-documented data accuracy issues over the years, acknowledged by Google itself more than once. That's not an argument against using any tool, bought, built, or straight from the source. It's the argument for never making a real decision off one snapshot from any of them. Trend over time, checked against more than one source, is what experience in this industry actually teaches — not which vendor to trust, but how rarely a single number from any of them deserves to be trusted alone.
Which is really the whole argument, said plainly rather than worked up to. None of this was ever about which tools you buy, or staying a step ahead of the competition on what gets measured. Buying the newest GEO tool on the market is irrelevant if what it hands back is junk, or if it's measuring something that was never going to help the client in the first place. The tool was never the point. Whether what you're measuring is actually true, and actually useful, always was.
Worth naming the scale of it too. Redefining how SEO gets evaluated and reported is a bigger decision for someone leading a department of two or three people than it is for a lone consultant working it out alone at a kitchen table. But if none of this actually needed changing, the industry wouldn't be full of people complaining about acronym overload, buzzword fog, and genuinely not knowing what half of this means anymore. That complaint is the evidence.
None of it comes down to the tool, in the end. You're not as good as your tool. You're only as good as the value you actually bring — and that was always the whole point.