The one thing AI cannot synthesise from somewhere else
Most GEO advice is about format — how to structure a page so an engine can lift it. That matters, but it misses a deeper lever. An engine can rephrase anyone’s explanation. It cannot produce a number that exists only in your systems. The most durable way to be cited is to hold a fact the synthesiser cannot synthesise.
the short answer
Generative engines synthesise what already exists — they cannot generate primary data. If you publish original numbers only you have (aggregated product telemetry, surveys, category benchmarks), an answer that needs that fact has nowhere else to go but to you. The 2024 GEO research found that adding statistics and source citations raised visibility in generative engines by up to ~30-40%, while keyword stuffing lowered it. Be the source of the fact, not one more voice with an opinion.
key takeaways
- Generative engines synthesise what already exists; they cannot produce primary data. A source holding a unique number is structurally hard to leave out when the query touches it.
- The 2024 GEO paper found that adding statistics, source citations and quotations raised visibility in generative engines by up to ~30-40%; keyword stuffing lowered it ~10%.
- Your own data — aggregated product telemetry, surveys, category benchmarks — is content no competitor and no model can replicate.
- For B2B SaaS: publish the number only your vantage point can produce, about your product, your customers in aggregate, or your category.
- It is not only format: citable substance — original data — is what turns a page into a source AI cites by necessity rather than by luck.
what raised visibility in the 2024 GEO study — and what hurt it
In words, so the bars do not carry it alone: the 2024 GEO study tested how specific content changes moved a page’s visibility inside generative engines, against a baseline. The methods that helped most were substantive — adding quotations, citing sources, and adding statistics each raised relative visibility by roughly a third to two-fifths. The one that hurt was the old keyword trick: stuffing lowered visibility by about a tenth. The lesson points away from manipulation and toward substance, and the most substantive thing you can add is a number nobody else has.
Why this matters more now, not less
It would be easy to read "publish original data" as timeless advice that predates AI, and in part it is — good research always earned links and trust. But synthesis changes the stakes in a specific way. When a human reader wanted an answer, they might read your essay and a competitor’s and decide; your opinion had standalone value because the reader did the synthesising. When an engine does the synthesising, your opinion is absorbed into a blended answer that may not name you at all, because a hundred other pages offered the same view. The essay’s scarcity collapses. What does not collapse is the proprietary fact, because synthesis cannot manufacture it. So the same shift that devalues generic commentary raises the value of original data — the two move in opposite directions, and the gap between them is widening as engines get better at the blending.
This reframes the content question many teams are stuck on. The anxious version is "how do we publish more to stay visible as AI absorbs our traffic." The better version is "what do we know that no one else can publish, and how do we make it impossible to answer certain questions without us." The first is a treadmill that synthesis will always outrun. The second is a moat, because it is built on something the engine structurally lacks. Moving from the first question to the second is most of the strategic work, and it usually means publishing less, not more — fewer pages, each carrying a fact that is genuinely yours.
The original-data advantage, in three parts
What an engine can and cannot generate, why a proprietary number is citable by structure rather than tactics, and where your original data is already sitting. Open each layer for the part that changes your plan.
01 What AI can and cannot generate
A generative engine is, at heart, a synthesiser: it reads across many existing sources and produces a fluent answer that blends them. That is enormously powerful for anything that has already been written down somewhere — definitions, explanations, summaries, opinions, re-reported facts. It is powerless for one thing: a primary fact that exists nowhere in its sources. A model cannot tell you what percentage of your customers did something last quarter, because that number lives only in your systems. It cannot benchmark your category from a vantage point it does not have. When a question reaches for a fact that only one source holds, the engine has no choice but to reach for that source — or to guess, which is the failure it most wants to avoid. The asymmetry is the whole opportunity: be the holder of the fact the synthesiser cannot synthesise.
02 Why original data is structurally citable
Most content competes in an ocean of substitutes. Your explainer on a topic sits beside a thousand others that say roughly the same thing, and an engine can assemble its answer from any of them, citing none in particular or whichever it trusts most. A proprietary number has no substitutes. If the only credible source for "how often B2B teams do X" is your survey, then an answer that needs that figure must come back to you, and the citation is not a matter of out-optimising rivals — there are no rivals for that specific fact. This is what "structurally citable" means: the citation is forced by scarcity rather than earned by tactics. The 2024 GEO research pointed the same direction from the content side, finding that adding statistics and source citations lifted visibility in generative engines markedly, while keyword stuffing actively hurt. Substance with numbers wins; padding loses.
03 Where your original data already lives
The objection is always "we do not have original research," and it is almost always wrong. Original data is not a thing you must go acquire; it is a by-product you already generate and have not packaged. It lives in three places for most companies. In your product: the aggregated, anonymised patterns of how customers actually use what you sell. In your customers: what they would tell you in a survey, a poll, or a set of structured conversations you could run this month. And in your category position: the benchmarks you can compute precisely because you sit across many participants and see averages no single one of them can. None of this requires a lab. It requires looking at the data you already own with a new question — which of these numbers answers something a buyer asks — and then doing the unglamorous work of aggregating it honestly and writing it down.
How to publish a number so it gets cited
Holding original data is half the work; publishing it so an engine can use it is the other half, and it is simpler than the technical framing suggests. The figure has to be liftable: stated in plain text as a clear sentence with its unit and context — "in our survey of 800 B2B marketers in early 2025, 62% reported X" — not buried inside an image or a chart with no text equivalent, where a model cannot read it. It has to be verifiable: you say how you got it, the sample, the period, the method, in a sentence a careful reader can trust, because an unsourced number is one an engine has reason to ignore. And it should have a stable home: a page that exists to host that figure, dated honestly, so the fact has a canonical address to be cited from.
Beyond that, the ordinary habits of citable content apply — a direct answer near the top, a heading that matches the question the number answers, a structure an engine can parse. But do not let technique crowd out the substance. A perfectly formatted page with no real fact is still just another voice; a plainly written page with a number only you have is a source. The 2024 research that found statistics and citations lifting visibility was measuring exactly this preference: engines reach for content that carries verifiable, specific substance, because that is the content that lets them answer without guessing. Give them the fact, say it plainly, and show your work.
The B2B angle: your category sees you as the source
For a B2B SaaS company this is not a side tactic; it is one of the strongest moves available, because your vantage point is genuinely unique. You sit across many customers in a specific category, which means you can compute things no individual participant can see: the real averages, the adoption curves, the patterns in how the work actually gets done. A buyer asking an engine about your space is asking a question your data can answer better than anyone’s, and an annual benchmark built from your aggregated, anonymised view becomes the kind of reference the whole category quotes — including the engines that now mediate the category’s questions. That is authority earned through fact rather than asserted through marketing, and it compounds: each year of data makes the next benchmark more valuable and harder for a newcomer to match.
The discipline is to choose the few numbers that matter and own them completely, rather than scattering shallow data across many posts. Pick the questions your buyers genuinely ask while deciding, find the figure from your own operations that answers each one, and publish it as a clear, dated, well-sourced reference you update on a rhythm. Over a few cycles you stop being one more vendor with opinions about the category and become the source the category’s facts come from — which is precisely what a synthesising engine reaches for, and precisely what a competitor cannot copy by writing a better blog post.
A warning: this only works if the data is honest
There is a sharp edge to this advice that is worth stating plainly. A proprietary number is a powerful citation magnet precisely because it is hard to check — and that is exactly what makes a dishonest one dangerous. If you inflate a figure, cherry-pick a flattering sample, or dress an opinion up as data, you may still get cited, and now an engine is repeating your distortion to buyers as fact, with your name on it. The same scarcity that makes original data citable also makes original error sticky: there is no competing source to correct it, because the whole point was that only you had it. A number that cannot be checked has to be one you would stand behind under scrutiny, because the day it is scrutinised the citation that helped you becomes the citation that exposes you.
So the discipline is method, not just publication. State the sample and the period. Describe how you collected and aggregated it. Be honest about what the number does and does not show, and about its limits. The rigour is not bureaucratic caution; it is what makes the figure trustworthy enough to be reached for in the first place, and durable enough to survive being right. Original data done carelessly is worse than no data, because it couples reach with error. Done honestly, it is the rare asset that is both widely cited and safe to be cited for.
The compounding asset: when a number becomes a series
A single original figure is useful; the same figure measured again, year after year, is a different order of asset. Once you publish a benchmark and repeat it on a rhythm, you stop owning a data point and start owning a trend — and a trend answers questions a single number cannot: not just "what is the rate" but "how is it changing," which is often the more valuable question and one almost no one else can answer about your category. The series also defends itself. A competitor who decides to copy your benchmark can start collecting today, but they cannot collect the three years you already have; the history is a head start that widens rather than closes, because each cycle you add extends the lead. That is the rare kind of content advantage that gets stronger the longer you hold it, instead of decaying when a rival publishes a better-optimised page.
Original research and AI citations: quick answers
Do I need a research team to do this?
No, and assuming you do is what stops most companies from starting. Original data does not mean an academic study; it means a number that is true, useful and yours, that nobody else can publish because nobody else has it. You almost certainly already generate that data in the normal course of business — what your customers do, what your product measures, what your category looks like from where you sit. The work is not running a research lab; it is noticing which of your everyday numbers would answer a question a buyer actually asks, aggregating it honestly, and publishing it clearly. A single well-framed statistic from your own operations, stated with its context and method, is worth more as a citation magnet than a long essay of opinion that an engine could assemble from a hundred other pages. Start with one number you already have.
What actually counts as original data?
Anything factual that originates with you and is not available elsewhere. The most common sources are three: aggregated product telemetry (patterns across how your customers use what you sell, anonymised and rolled up), direct research (a survey of your audience, a poll, a structured set of interviews turned into figures), and category benchmarks (numbers you can compute because you sit across many customers in a space and can see the averages no single participant can). What unites them is exclusivity: the figure exists because of your vantage point, so a model answering a question in your space has nowhere else to get it. Opinion, summary and re-reported third-party statistics do not count — an engine can synthesise those from everyone. The test is simple: if a competitor could publish the same number by reading the same public sources, it is not original, and it will not make you uniquely citable.
Isn’t this just content marketing with a new label?
It overlaps with content marketing but the emphasis is different, and the difference is the point. Classic content marketing often rewards volume and opinion — publish frequently, have takes, build an audience. The original-data approach rewards scarcity and fact: publish less, but publish numbers only you can produce. In a pre-AI world the two looked similar because human readers valued both a good essay and a good statistic. In a world where an engine assembles the essay for free from everyone’s takes, the essay loses its scarcity and the proprietary number keeps it. So this is content marketing reweighted for how synthesis works: away from being one more voice with an opinion, toward being the only source for a fact. The brands that thrive are not the ones publishing the most; they are the ones publishing what cannot be published by anyone else.
How do I make my data easy for AI to cite?
Make the number liftable and verifiable. Liftable means the key figure sits in plain text, stated as a clear sentence with its unit and context — "in our 2025 survey of 800 B2B marketers, 62% reported X" — not locked inside an image, a chart with no text equivalent, or a vague paragraph an engine has to interpret. Verifiable means you state how you got it: the sample, the period, the method, in a sentence a careful reader (or a model) can trust. Give the figure a stable page that exists to host it, date it, and keep that date honest. Beyond that, the same structural habits that help any citation apply — a clear answer near the top, a plain heading that matches the question the number answers. But the foundation is simpler than technique: have a real number, say it plainly, and show your work. Engines cite what they can quote with confidence.
A note on sources and certainty
The figures here come from the 2024 GEO study (Aggarwal and colleagues, "Generative Engine Optimization"), which measured how content changes affect visibility inside generative engines and found substantive additions — statistics, source citations, quotations — lifting it by roughly a third to two-fifths while keyword stuffing reduced it. We treat those as directional evidence that engines favour verifiable substance, not as exact multipliers for your pages; the precise numbers depend on the engine, the query and the content, and they will move as systems change. The deeper claim does not depend on any single figure: a generative engine cannot produce a primary fact it has no source for, so holding such a fact is a structural advantage no amount of competitor optimisation can erode. The AC Group has spent 27 years arguing that authority is built on what others can verify about you — and original data is the purest form of that argument, because it is verifiably, exclusively yours.