Somewhere in your market this week, an assistant recommended a competitor to a buyer who never touched a search results page. You did not see it happen. There is no dashboard where that shows up, no alert, no line item. The only reason you would find out is if that buyer mentioned it on a call.
This surface did not exist when your site was built, so nobody on your team is behind on it. There is no established practice to catch up to, no competitor who quietly solved this two years ago while you were not looking. That is unusually true here, and it is the actual reason this is worth ten minutes of your attention.
The scale behind that invisibility is bigger than most dashboards admit. In 2025, Gartner surveyed 645 B2B buyers and found 45% had already used generative AI to research vendors before contacting a rep[1]. Most of that research produces no click at all. A 2025 Pew Research study of nearly 69,000 searches found people clicked a result on just 8% of searches that showed an AI summary, and clicked a link inside the summary itself on only 1%[2]. In 2024, before AI summaries were common, zero-click was already the default at 58.5% of US Google searches[3]. Google’s own 2026 update gave GA4 the AI referral traffic most dashboards still miss[4], and even that only catches sessions that still carry a referrer. An assistant that answers and gets nobody to click is invisible by design, not by a gap somebody forgot to close.
What follows is not a system. It is one traced client acquisition, what has since been verified about it, what has not, and what changed here as a result. That is a smaller claim than most content on this topic makes, and it is the only version compatible with actually being honest about it.
What Happened, Precisely
One client relationship on record here traces directly to an AI assistant. A Series B compliance technology company found this practice because ChatGPT quoted a paragraph from a blog post, and the prospect called within days. That is one traced instance, not a repeatable system, and treating it as more would be dishonest.
Here is what is actually known. The company builds privacy and compliance software, several years into a funded raise, with no prior relationship to this practice. Someone on their team asked an assistant a question shaped roughly like “who fixes a marketing website nobody technically owns,” and the assistant’s answer named this practice, apparently quoting a line from a blog post that was already live. The prospect called days later, already familiar with the argument, and the deal closed faster than any other client relationship on record.
Here is what is not known, and it matters more than the part that is. Nobody can say with certainty which page the assistant actually pulled from. It could have been the blog post. It could have been a LinkedIn profile, a mention somewhere else entirely, or the model simply recalling a name it had already seen associated with the topic during training. Nobody thought to ask for a screenshot of the exchange at the time, and by the time it seemed worth asking, the exact wording was gone. One traced acquisition, with a genuine gap in the middle of the trace.
That gap is the reason this whole piece is written the way it is. It would be easy to round “an AI assistant surfaced me to my best client” into “here is the system that gets you cited.” The honest version is smaller and less exciting: one instance, verified as far as it can be verified, sitting next to a year of deliberate follow-up work that has not produced a confirmed second one yet.
Three Things That Showed Up Consistently
Three patterns showed up across the traced acquisition and a year of deliberate follow-up testing: plain pages that state what a company does, who it serves and what it costs; a description that reads the same on every surface where the company appears; and presence on third-party pages an assistant already trusts more than a company’s own site.
Plain pages won out over persuasive ones. The pages that get lifted into an assistant’s answer are the ones that state, in a sentence or two, what a company does, who it is for, and roughly what it costs, in language that can be quoted without interpretation. A homepage built around a clever narrative line, with no category words anywhere in the headline, tests poorly on this measure even when a human reader likes it more.
Consistency mattered almost as much as the content itself. The same company, described three different ways across a homepage, a professional profile and a directory listing, reads to an assistant as a disambiguation problem rather than three sources agreeing. The failure is rarely that any one description is wrong. It is that none of them match closely enough for a model to be confident it is looking at one entity instead of several.
Third-party presence carried real weight, more than link-building habits usually assume. In 2025, Ahrefs studied 75,000 brands and found branded web mentions correlated with AI Overview visibility at 0.664, more than double the 0.218 correlation for raw backlink count[5]. Being talked about on someone else’s page outweighs being linked to from it, which is a different skill than most SEO checklists teach.
The consistency pattern showed up on a smaller scale during the follow-up work here too. Two profiles on the same professional network, both technically real, sat inside this site’s own entity markup as though they were interchangeable. No study was needed to know that was a problem the moment it was noticed. Fixing it was half a day of unglamorous cleanup, not a redesign, and it deserves its own dedicated page rather than a paragraph here.
What I Could Not Verify
The evidence here does not support two common claims: that schema causes AI citations, and that a single position check means anything. A page with zero structured data was cited twice in one answer, while this site’s fully populated schema graph was cited zero times. The same prompt, run twice 90 seconds apart, flipped from a top citation to none.
The schema comparison came from testing this site’s own homepage against a competitor’s page targeting the same buyer question. That competitor’s page carries no JSON-LD schema at all, and an assistant cited it twice in a single answer. This site carries a fully populated schema graph across seven service categories, several structured-data types deep, and was cited zero times in the same test. On that evidence, schema richness does not predict who gets cited.
In 2026, Ahrefs ran a more rigorous version of the same question. It compared 1,885 pages that added JSON-LD schema between August 2025 and March 2026 against roughly 4,000 pages that did not. Citations in Google’s AI Mode moved 2.4% and ChatGPT citations moved 2.2%, both inside statistical noise. Citations in Google AI Overviews actually fell 4.6%, a result Ahrefs called significant[6]. Schema still earns rich results and gives search engines a cleaner read of who a company is. It is not, on the evidence available anywhere, a citation lever, and that includes schema’s actual job on a B2B SaaS marketing site.
The single-run problem is worse than most people testing this expect. An identical prompt, run through the same model with web search on, twice, 90 seconds apart, produced a top citation on the first run and no mention at all on the second. Nothing about the site changed in that window.
The only honest conclusion is that a single check tells you almost nothing about your actual position, and anyone reporting “we are cited number one by ChatGPT” off one run is reporting noise, not a result. Semrush’s 2026 analysis of 126 million AI search prompts found a similar instability one layer up: the overlap between brands an assistant mentions in an answer and brands it actually cites with a link falls as low as 30% on Gemini[7]. Being named and being cited are not reliably the same event inside a single answer, let alone across two runs of the same prompt.
How to Check Your Own Position Without Fooling Yourself
Checking your AI citation position with one prompt on one model produces a number that means nothing, because the flip behavior above is not an edge case. The honest version runs the same locked set of prompts through the same models multiple times and reports a hit rate, citations out of total runs, never a single yes or no.
The instrument matters more than the question you ask it. A panel of prompts reworded between checks, or run through a different model each time, produces numbers that cannot be compared to each other. None of that variation would be about your site. All of it would be about the instrument you used to measure it.
The version used for the follow-up work here runs each prompt three times per model, across a fixed panel of twelve buyer-shaped questions, and never changes the wording between sweeps. Two models, both tested with live web search enabled. The result recorded for each run is not just whether the site got cited, but which specific URL got cited, because that is what actually explains a miss instead of just reporting one.
None of this is complicated. It is tedious, and tedious is exactly why almost nobody does it properly. Most of what passes for “we’re tracking AI citations” on a marketing team is one person typing a question into ChatGPT once a week and reporting whatever they see that day. That is not measurement. It is a mood check.
What This Is Worth Doing For, and What It Is Not
This work is worth doing because it is cheap, the pages that win are ones any B2B SaaS marketing team can already build, and almost nobody in the category does it properly yet. It is not a pipeline strategy: one traced client, however real, is a data point, not a forecast, and nobody honest promises it happens twice.
The category is genuinely empty right now, which is rare and worth taking seriously. Most competitors in this space are not writing plain, quotable pages, are not checking whether their own descriptions agree with each other, and are not measuring citation position with anything more rigorous than an occasional manual check. Doing the basic version of all three moves a company from invisible to plausible on a surface almost nobody else is optimizing for on purpose, which is a genuinely different situation from most of the split between getting a page reached and getting it cited, where the reachability half is at least being worked on by somebody.
What it is not, and should not be sold as, is a substitute for the channels that actually produce volume. One assistant recommendation produced one real client relationship over roughly a year of deliberate work. That is a genuinely good return on a small amount of effort. It is not a forecast for next quarter’s pipeline, and treating it like one would be the same mistake as reading three clicks on a blog post as a trend, which is the same blind spot that shows up in analytics dashboards every day.
There is no conversion number attached to any of this, and there should not be. The pages get built, the description gets made consistent, the position gets checked honestly instead of once. What a buyer does after an assistant mentions a company’s name is the marketing lead’s number to own, not a number this kind of work can promise in advance.
Sources
- Gartner, 69% of B2B Buyers Turn to Sales Reps to Validate AI-Generated Insights – Survey of 645 B2B buyers fielded Aug-Sept 2025; 45% used generative AI to research vendors and products before contacting a rep ↩
- Pew Research Center, Google users are less likely to click when an AI summary appears – 68,879 searches from 900 US adults tracked March 2025; 8% of searches with an AI summary produced a result click vs 15% without ↩
- SparkToro + Datos, 2024 Zero-Click Search Study – Datos clickstream panel, tens of millions of panelists, Sept 2022-May 2024; 58.5% of US Google searches ended in zero clicks ↩
- Google Analytics Help, GA4 Release Notes – May 13, 2026 entry; native AI Assistants channel added to the GA4 Default Channel Group report for referrer-carrying sessions; does not cover zero-click answers ↩
- Ahrefs, What Correlates With AI Overview Brand Visibility – May 2025; 75,000 brands; branded web mentions correlate at 0.664 vs 0.218 for raw backlink count ↩
- Ahrefs, We Tracked 1,885 Pages Adding Schema. AI Citations Barely Moved. – 2026; 1,885 pages that added JSON-LD vs ~4,000 controls; ChatGPT/AI Mode changes within noise, AI Overviews citations down 4.6% (significant) ↩
- Semrush, Expanded 2026 AI Visibility Index – 126 million US AI prompts, Jan-Apr 2026; brand-mention-to-citation overlap as low as 30% on Gemini ↩
Seeing these patterns at your company?
Book a free WebOps Diagnostic. I'll review your site before the call and share specific observations.
Book a Free Call →Frequently Asked Questions
The evidence says no. In 2026, Ahrefs tested 1,885 pages that added JSON-LD schema against roughly 4,000 that did not, and found citation changes in ChatGPT and Google AI Mode inside statistical noise, with AI Overviews actually down 4.6%. Schema still earns rich results and cleaner entity data. It does not earn citations.
No. An identical prompt, run through the same model with web search on, twice, 90 seconds apart, produced a top citation on the first run and no mention at all on the second. A single check is noise, not a result. Measuring this honestly means running the same locked prompt panel multiple times and reporting a hit rate.
Plain pages that state what a company does, who it serves and what it costs, in language that can be quoted without interpretation, described the same way on every surface where the company appears. Persuasive homepages built around a clever narrative line with no category words in the headline test poorly on this measure, even when readers prefer them.
It is worth doing because it is cheap and almost nobody in most B2B SaaS categories has done the basics yet. It is not a pipeline strategy. One traced client arrival through an AI answer is a real data point, not a forecast, and nobody honest can promise it happens on a schedule.
Run a fixed panel of prompts through the same models at least three times each, without changing the wording between checks, and record a hit rate rather than a single yes or no. A one-time manual check into ChatGPT tells you almost nothing, because the same prompt can flip from a citation to none within minutes.
