All research
The serious research on AI Search is really only about 18 months old. In that time it has produced a remarkable amount of data and a remarkable amount of disagreement.
Those overlaps are the closest thing this field currently has to settled knowledge.
The problem we've seen is that these studies count different things, in different industries, on different kinds of questions, and almost nobody has put them side by side. Which means you can cherry-pick a number without knowing whether it holds up anywhere else.
So we read 23 AI Search studies properly, including one of our own.
We checked each one for sample size, method, dates and which kinds of company it covered, then looked for the places where independent studies using different methods landed in the same spot. Including our own research.
Those overlaps are the closest thing this field currently has to settled knowledge.
Six studies we looked at have measured where AI gets its information about brands. The results run from 1% to 85%, depending on whether the number counts your own website or everything except it, whether paid content is included, and whether the question that produced the answer named your brand at all.
They aren’t contradicting each other. Each one counted something different. See what we mean below:
Line up the definitions and the disagreement mostly disappears. AirOps puts it most directly: when AI mentions a brand, 85% of the sources it cites are third-party, not the brand's own site. A brand is 6.5 times more likely to be surfaced through someone else's content than its own.
Ranqo tracked 149,912 citations across five AI engines over three months. The company being asked about accounted for 2.9% of them. McKinsey found roughly 1% in consumer goods.
Omniscient analyzed 23,387 citations from queries that named a specific brand and found the brand's own site supplying 23%. The other 77% split between earned media at 48% and content published by other companies at 30%.
The picture changes with the question. When people asked whether a specific product they named in the prompt did a particular thing, owned content supplied half the citations. When they asked what customers thought, earned media supplied 82%.
It changes with the engine too. AirOps found first-party citations run 13% to 21% on Claude and Perplexity, against 4% to 11% on GPT-5.
The pattern is straightforward: if you ask AI which tools are good for something, it barely touches company websites. Ask whether a specific product does a specific thing and it goes looking for details, which is where your own pages carry weight.
For the question most buyers start with, the one where they don't yet know your name, your website is close to invisible.
Ranqo's 149,912 citations split three ways. The brand's own domain took 2.9% of citations. Other companies' pages took 75.2%, meaning competitors, peers and vendors selling to the same buyers. Everything else, video, media, forums, reference and review sites, made up the remaining 21.9%.
So bucket every citation that isn't your own website, whether it belongs to another company or to a publisher or a forum, and you get 97.1%.
To put Ranqo's split concretely, if an AI answering a question about your category looks at twenty sites, roughly fifteen are company websites and at most one is yours.
The company-website share also swings hard by engine and by industry. Gemini draws 69% of its citations from company sites, ChatGPT 47%. Cybersecurity runs at 74%, government and nonprofit at 16%.
Profound’s sample is far bigger: 11.8 billion citations, eight models, 29 industries. On your own site it lands close to Ranqo, at 2 to 6%. It splits the rest differently, putting all corporate pages at 57% against Ranqo’s 78%.
What the studies agree on is this: most of what AI cites is company websites. It just isn’t your company website.
The usual conclusion from all this is that you need press coverage, but the numbers don't support it.
Muck Rack found that about 27% of links cited by AI are journalistic, and that the figure has held between 20% and 30% across every study they've run. That comes from a PR company, which makes it worth taking seriously. Profound found software companies get 11.4% of their citations from earned media, the lowest of all 29 industries measured. Ranqo put news and business media at 3.8%.
A journalistic feature in a national title is certainly worth having for other reasons, but it is not much of what AI reads.
AirOps found that roughly 90% of third-party brand mentions came from listicles, comparison pages and roundups. Ranqo found the same from the other direction: the ranked "best X" article is the single most-cited type of page, at about 21% of all citations.
There's a reason one format dominates. A single "best project management tools" article gets pulled into dozens of different AI answers, because it already contains the shape of the answer AI is trying to produce. Position inside it matters too. AirOps found 80% of mentioned brands appear in the first three slots of the page that cites them.
Another tidbit we found: Reddit turns up in roughly one in five AI answers and supplies about 3% of citations. Sounds contradictory, but both are true. Frequency and share are different measurements, and a lot of published advice confuses them.
Run your ten most important buying questions through ChatGPT, Gemini and Perplexity, and write down every page each answer draws from.
On each one, check two things: whether you’re on it, and whether what it says about you is still true.
That list is your outreach target. It will look almost nothing like a media list, and a good share may even belong to your competitors. Find the contact for each page and make the case for your brand.
Ask ChatGPT the same question ten times and you’ll get ten different lists. Plenty of people have seen that and concluded AI visibility can’t be measured.
The observation is right, but the conclusion doesn’t follow, because two things are moving here.
Whether your brand gets named at all is close to fixed. Run the same question on the same engine over and over and you’re usually either in most of the answers or in almost none of them. That describes where you stand today, by the way, it doesn’t mean you’re stuck there.
Where you sit in the list is what keeps moving: second one run, fifth the next. Fishkin’s read after running the experiment was that tracking the order is a waste of time, and tracking whether you appear is worth doing.
SparkToro ran the experiment in January 2026. 600 volunteers put 12 prompts through ChatGPT, Claude and Google's AI a combined 2,961 times, each prompt run 60 to 100 times per tool. Volunteers used their own accounts, with whatever history and custom settings they already had, so the results show what ordinary people actually get.
Ask an AI tool for recommendations a hundred times and there's less than a 1 in 100 chance any two responses give the same list of brands. For the same list in the same order, it's closer to 1 in 1,000.
One cancer hospital showed up in 69 of 71 ChatGPT responses about the best care on the West Coast. It was the top recommendation in 25 of them. Almost always present, almost never in the same place.
Rand Fishkin ran this research expecting to show that AI tracking was a waste of money. His own data changed his mind. Tracking the order is pointless, he concluded, but tracking whether you appear at all is worth doing. He published that even though it undercut the argument he set out to make.
Running one prompt a hundred times tells you how the model varies. It doesn't tell you much about real use, because real people don't ask the same way twice.
SparkToro tested that as well. They asked 142 volunteers to write their own prompt for finding headphones for a traveling family member. Semantic similarity across those prompts came out at 0.081, which is close to no overlap at all. One intent, expressed 142 different ways.
Those prompts produced 994 responses. Bose, Sony, Sennheiser and Apple appeared in 55% to 77% of them.
The wording changed completely and the same brands kept surfacing. That tells you phrasing isn’t the lever. It also tells you what you earn is durable: presence that survives 142 different ways of asking is presence you don’t have to re-win every time someone words the question differently. Fishkin is careful to say real-world variety is probably wider than any controlled experiment can capture, and he’s right to be.
Ranqo measured the same behavior across a much bigger sample: 102 brands, 102,025 responses, five engines, three months.
For every combination of brand, question and engine tracked over at least three runs, they classified the pattern. 63.2% were never mentioned. 14.3% were always mentioned. Only 6.8% genuinely flipped back and forth.
That leaves 77.5% of cases sitting at one extreme or the other, and 93.2% outside the volatile middle. Ranqo's data shows that absence holds steady without saying why it holds.
The likeliest reason is the simplest one: nothing in the sources AI reads connects your name to that question yet, so there’s nothing there for the model to draw on. That’s the part you can act on.
Presence in the AI answer barely drifts on its own, so when it does move, that's a signal.
Fractl ran 96 prompts 15 times each through GPT-4o, Gemini 2.5 Flash and Claude Sonnet 4.6, producing 4,320 responses and more than 8,500 brand references.
Only 11% of brands appeared in all three models. 77% appeared in just one. Notion was cited 17 times more often by Claude than by Gemini. AirOps found the same shape in a separate dataset, with 68% of brands appearing on a single platform.
Ahrefs looked at 75,000 brands and found the opposite-looking result. The three AI surfaces it tested overlap heavily on which brands they mention, correlating between 0.749 and 0.821. Their summary was that the engines are still mostly mentioning Nike, Apple and Amazon.
Both are right. The engines agree on the household names and diverge sharply on everyone else, which is most brands.
So a dashboard reporting that your AI visibility went up should prompt one question: on which engine? A flat number can hide the fact that you're doing fine on ChatGPT and nowhere on Gemini.
Ranqo measured how often the framing of a brand changes between runs, against how often the mention itself changes. Sentiment shifted in 45.5% of cases. The mention shifted in 6.8%.
Sentiment is 6.7 times noisier, which means the two figures sitting side by side on your dashboard need very different amounts of data behind them. Ranqo uses a floor of ten prompts per platform per brand before it trusts a sentiment score.
A citation and brand visibility are two different things. A citation is a source AI draws from when it builds an answer. Visibility is your brand getting named in the answer itself. Semrush data across 1,094 categories found the most-cited domain and the most-named brand were the same only 20.8% of the time, and the two correlate slightly negatively at −0.229. The most-named brand had usually been cited at least once, 69.9% of the time, so they’re connected.
Being the domain AI quotes and being the brand AI names are different outcomes, and getting named is the one your buyers see.
A blended number can hide that you’re fine on ChatGPT and absent on Gemini. Those are two different problems with different sources behind them.
Where you never appear on an engine, you aren’t in the sources its answers get built from, so getting into them is the job. Where you appear sometimes, you’re on some of the pages carrying that answer and missing from others. That gives you two moves: get onto the pages you’re missing from, and on the pages that already carry you, ask the publisher to move you higher in the piece, since the brands AI cites tend to sit near the top of the pages that mention them.
AI gets facts about companies wrong at a measurable rate, and no platform offers a way to report it. That leaves one lever: editing the sources AI is citing in the first place.
Let’s start with how much AI actually talks when it answers us. In most cases, an answer engine adds context, comparisons, caveats and recommendations nobody asked for.
Profound analyzed 50,000 prompts across seven industries and five prompt types, and found that roughly 47% of the content in AI responses was unsolicited. Rationale for the answer, side-by-side comparison tables, fee and term caveats, rankings, shopping guidance, follow-up suggestions, etc.
Responses are long enough to need filling. ChatGPT averaged 3,043 characters per answer, Gemini 2,921, Claude 1,590. Profound calls this the Parrot Problem, because the models can only repeat what already exists about you.
So there is roughly twice as much being said about you as the question called for. And a measurable share of it is wrong.
Profound's FactCheck research checked more than 158,000 claims AI made about brands against those brands' own records. Between 4.3% and 6.8% were wrong, depending on the engine.
Buyers notice. G2 surveyed 1,076 B2B decision-makers in March 2026 and found 64% run into AI getting something wrong at least a few times a month.
The errors split almost evenly. 54% of brands had an inaccurate claim where AI was citing their own content, while 51% had one citing third-party coverage.
In this case, AI didn’t invent false information. It read the page and repeated what it found.
The problem is that you can edit your own pages in a matter of days or weeks. But you can’t go and edit the other half of sites you don't own, written by people who don’t even know there’s anything to update.
If the pages are the only thing you can change, it's worth knowing how AI weighs their age…and if playing a waiting game is an actual strategy.
Ahrefs looked at nearly 17 million cited URLs and found the average one is 1,064 days old, close to three years. Google's AI Overviews cite content older than what ranks in normal search. In short: old pages keep getting read.
A different 2025 study from Waseda University ran an experiment that suggests AI favors fresher-content. Researchers took passages, changed nothing except the publication date attached to them, and measured what happened across seven models. Passages labeled more recent were systematically promoted. The average publication year of the top ten shifted forward by as much as 4.78 years, individual passages moved up to 95 places, and where two passages were equally relevant, the fresher-looking date flipped the preference up to 25% of the time.
That being said, the authors treat this as a defect, so restamping a date isn't a viable strategic move. And nobody has tested whether publishing new material - or what looks to be new material - displaces the old. What the evidence does support is narrower and more useful: update the page that's already being cited. A corrected version of a page AI is reading carries a recency signal the models demonstrably respond to, on a page they've already chosen.
Check each one against what you actually charge today, then swap vague claims for retrievable ones. Models quote specifics: a number, a named integration, a certification. Phrases like "enterprise grade" leave them nothing to work with. Repeat it quarterly at minimum, and every time pricing changes.
Then, check every factual claim in those answers against your records. Follow each wrong claim back to the page AI cited, and build your outreach list. Start with whichever pages keep appearing, and ask the publisher for the update.
Two benchmark studies have tested whether editing your own pages changes how AI uses them, about a year apart. They agree on three things, and the three together say where owned-page effort pays and where it doesn't.
In August 2024 a team at Princeton published GEO: Generative Engine Optimization, built on 10,000 real search queries each paired with the top five Google results, with every answer generated by GPT-3.5-turbo. In June 2025 a group from Parameter Lab with TU Darmstadt, Mannheim, Tübingen and NAVER AI Lab published C-SEO Bench: ten published methods, six domains, four models, over 16,000 documents, and a protocol for what happens when many companies adopt the same method at once.
A persuasive, authoritative tone does nothing. Princeton found no significant improvement from rewriting pages to sound more authoritative, and read it as the engines already being robust to that kind of writing. C-SEO Bench tested the same thing on four models and found no significant gain on any of them.
Getting read early beats anything you write on the page. AI doesn’t read the entire web when it answers a question. It pulls a handful of pages, in order, and writes from those. C-SEO Bench measured what the order is worth: they changed nothing except which document the model saw first, and that beat all ten content methods they tested, in every domain and on all four models.
Princeton’s test shows where the order comes from. They gave the model the top five Google results for each question, so Google decided which pages made the handful and which got read first. When they applied the same content edits to all five, the pages that improved inside the answer had started third or fifth. The one sitting first lost ground.
So two things decide whether your page shapes the answer, and both are settled before the model reads a word of your copy: whether you’re in that handful, and how early you’re read. That’s retrieval, which is why search position still feeds AI visibility.
The gain shrinks as everyone adopts it. Princeton reran their experiment with every source optimized simultaneously instead of one. C-SEO Bench did the same and the result was near-zero gains for any company, which they call a congested, zero-sum game.
The more useful explanation is in Table 2 of the Princeton paper.
When they reran the experiment with every source optimized at once, the effect split by where the page already sat in Google's results. Everything in the table is a change in how the page fared inside the AI answer. Google rank is the condition, not the outcome:
Their caption says it plainly: these methods are especially helpful for websites ranked low in search.
So both results are right, and what's being helped is AI visibility rather than Google ranking. Princeton didn't test whether these edits move a page in Google, and says so in its limitations.
What the table shows is a catch-up mechanism inside the AI answer. If Google already puts you first, and the model therefore reads you first, adding citations or statistics costs you ground. If Google puts you fifth, the same edits win some back.
Both benchmarks put documents in front of a model and measured what happened next. Neither asked how much of what AI reads is company-owned in the first place. That share runs from 42% in one industry down to 6% in another, and it decides whether your own pages are a lever for you at all.
Bluefish scored 194 large brands across nine verticals in May 2026 and recorded which domains AI cited when discussing each brand.
Telecom was the outlier, and the best case for owned content anywhere in the study. 42% of citations came from company websites, four of the top ten source domains were company-owned, and verizon.com was cited third.
The other 58% is news and publishing, comparison and affiliate sites, social and e-commerce. Reddit is the most-cited source in telecom and CNET is second, both ahead of any company site. So owned content does more work in telecom than in any vertical Bluefish measured, and it is still the minority of what AI reads.
The other verticals Bluefish published in full go the other way. In pharma, company websites are 6% of citations and none of them reaches the top ten, where trade press and news publishing supply 51%. In credit cards, company websites are 25%, and the one that makes the top ten sits tenth, behind comparison and personal finance sites.
The split tracks what AI treats as the decisive question in each industry. Telecom questions are about network performance, plan structure and device management, which are specifications the operator is the only real source for. Pharma questions are about product safety, clinical integrity and supply chain, which are trust questions where AI defers to trade press and regulators.
One caveat here: this is evidence about what AI reads once your brand is already in the conversation. What gets you into an answer when nobody named you is a different question. This mirrors what we've already seen: third-party pages dominate the citation list across most industries. Even, in this case, for branded prompts.
Pull the top twenty sources cited for your category's buying questions, with nobody's brand name in the prompt, and count how many are company websites. If it's a real share, your own pages are a primary lever. If it's close to none, they aren't, and no amount of publishing will change that. Run it again with your brand named: the two lists usually differ, and the unbranded one is where discovery happens.
Add citations and real statistics where they fit your buyers' questions. This won't move you in Google, and nobody has tested whether it would. What the evidence supports is a better showing inside AI answers for pages outside Google's top three, and the effect is larger in factual, legal and opinion-led categories.
The same three measurably cost ground in AI answers for pages already ranked first. What's left on your own site is keeping the facts right and the page up to date.
Pull up your AI referral traffic. For most companies it's insignificant, a few hundred sessions against hundreds of thousands. But if you were to ask your newest customers where they first came across you, some will say ChatGPT or AI Search.
Both are true. AI recommendations produce real demand, and almost none of it arrives as a click you can trace.
Scrunch ran an opt-in panel between February and May 2026 that linked people's real AI conversations to the same people's web activity afterwards: Google searches, brand site visits, retailer product pages.
When an AI platform recommended a brand to someone with no recent sign of using it, over the following week that person became roughly:
There's an obvious objection to all of this. Someone who asks ChatGPT about running shoes was probably already shopping for running shoes, so did the answer move them, or did it just catch them mid-purchase?
Two things in the design speak to that. Each person was compared against their own earlier weeks, so somebody who simply had a busy week doesn't register as a lift. And Scrunch ran the same analysis on people who were already customers of the brand, where the mentions made no measurable difference.
The study's findings show that being recommended does roughly double the work of being mentioned in passing. And position inside the answer matters: brands named first were 389% more likely to be searched afterwards, brands named last only 50%.
Similarweb's The Downstream Impact of AI Visibility tracked an opted-in panel of US desktop users across finance, travel and beauty between July and December 2025. They isolated people who asked ChatGPT something, got a brand recommendation back, hadn't named that brand in their prompt, and hadn't visited its site in the previous four weeks. Then they watched the next seven days.
Those users were 2.5 times more likely to visit the recommended brand's site.
The part that matters for your reporting is how they got there. 55.9% of AI-influenced visits arrived via search, against 40.4% of ordinary visits. Direct ran at 19.9% against 38.8%. Visible AI referrals accounted for 8.8%.
Someone reads an AI answer, remembers the name, and types it into Google days later. Your reporting logs that as organic search.
Those visitors are also worth more than average. Similarweb found AI-influenced visits ran to roughly twice the pages and twice the time on site.
The first is the one above: demand that AI created arrives labeled as organic or direct.
The second is cruder. Search Engine Land tracked more than 50,000 AI Overview events to a single brand over nine months and found 22.4% were recorded as Direct rather than as search at all. So even the trackable portion gets misfiled.
One study has measured the size of the gap. Graphite compared n8n's analytics against a survey of their own converting customers. Last-touch attribution credited AI with about 1% of conversions, but the survey put it at about 9%. That's one company in one category, so treat it as an illustration rather than a rate, but the direction matches everything above.
As Scrunch puts it: most people still don't act in any given week. So the value here is in reaching new people at the top of the funnel, repeated often, rather than a steady conversion rate.
Stop asking your AI visibility vendor for a referral traffic number and watch branded search volume instead, and watch it against your visibility in the engines your buyers use. You should expect a lag. Both panels here use a seven-day window, so movement follows a placement by days, or even weeks, not hours.
Ask new customers where they first came across you, in a required “How did you hear about us?” section of an onboarding or demo-booking flow. Compare that answer against what your analytics said about the same cohort. The gap is your real AI number.
Competitors can bid on the demand your AI presence generated, and Similarweb's own read is that this is already pushing branded CPCs up. A visitor who searches your name after an AI recommendation still has to find you at the top of the results.
The demand is real. Every engine recognizes your company when someone names it directly, and most of them stay quiet about you when someone asks your category question instead.
How well known you already are is a factor. About one brand in eleven breaks the pattern, roughly half in a direction you can fix, and none of the fixes have to do with your own website.
Ranqo sorted 102 brands into three groups by how well known they are, then measured how often each turned up when someone asked a category question without naming any company.
Global household names, like Stripe and Nike, appeared in 72.9% of those answers on their very first tracking run. Established mid-market and regional brands appeared in 43.6%. Small and niche brands appeared in 11.4%.
The three groups held 11, 36 and 55 brands. Ranqo sorted them using Wikipedia presence, recent coverage in major US and UK outlets and company size, and flags the eleven-brand top tier as too small to carry strong claims of its own.
This moves gradually rather than all at once. The fastest-moving small brands in Ranqo's cohort gained 10 to 20 points over three months. Set against the 11.4% their tier averages, that is a large gain, and it took a quarter to earn.
Ranqo is careful about what this proves, and so are we. They sorted the brands by things like Wikipedia length and press coverage, which are already signs of being well known, so the ladder measures the size of the gap rather than explaining it. Whether getting better known is what makes AI start naming you is a question this study doesn't answer.
What it does show is what the brands AI names most have in common: a Wikipedia article, recent coverage in Tier-1 outlets, a steady presence on third-party pages and social platforms. A solid baseline of coverage you don't own is what the top tier has and the bottom tier doesn't, and it is the part of that gap you can go and build.
Ranqo tested the other half of this directly. When the prompt contained a company's name, the answer mentioned that company between 94% and 100% of the time, on every engine, from the first run. Ranqo is upfront that it's a low bar: if the question hands over the name, the answer hands it back.
Put the two together and you have the problem in one line. Ask AI about your company and it answers almost every time. Ask your category's buying question without naming anyone and a small brand shows up about one time in nine. The model is not short of information about you. What's missing is anything tying you to the question your buyers actually ask, and that tie gets made on other people's pages.
Kevin Indig analyzed Semrush's AI Visibility data across 1,094 US categories, five prompts each, in monthly snapshots from January to June 2026 on ChatGPT.
Some categories have a brand AI reaches for almost every time. Indig calls that brand the category's owner, and he sets the bar at appearing in four of the five prompts and finishing at least five points clear of whoever is second.
By June, only 15.2% of the 1,094 categories had one. In 53.7%, several brands were still trading places between snapshots, with no single name AI reliably reached for.
The busier the category, the less likely it was to have an owner: 11.3% in the higher-demand half against 19.0% in the quieter half. Add up the AI search demand behind all 1,094 and 89.3% of it sits in categories nobody owns.
That bears directly on the most common version of this problem: if AI knows your name but never brings you up in your category, the first question is whether that category already has an owner. Weighted by demand, nearly nine times in ten it doesn't.
Fractl mapped brand references against each brand's traditional search footprint. More than nine in ten lined up: stronger search authority generally meant stronger AI visibility. The interesting part were the rest of the brands.
Fractl found 377 brands that the models referenced far more often than their search footprint predicted, and every one of them had the same thing going for it: they appeared again and again on pages other companies published. Roundups, expert lists, comparison reviews.
Monday.com was the largest overperformer in the set. Six of the top fifteen were education brands, among them Stanford, MIT and Khan Academy, and Fractl reads that as credibility doing the work rather than domain rating.
So if a competitor with half your organic traffic appears in the answer and you don't, go and look at the coverage they've earned.
Fractl's data compares different brands at one moment. What that design can't show is what happens to a single brand before and after, which is the question we went looking into.
Noble tracked 121 brands over 18 weeks, from January 26 to May 25, 2026. Of those, 79 earned at least one of the 599 brand mentions that went live in third-party sources, and 41 earned none and served as a control group.
The clearest group was 27 brands that began essentially invisible, under 1% visibility, and went on to earn four or more mentions. Within four weeks of their first mention going live, they moved from roughly zero to 19%. From absent, to appearing in about one in five answers, inside a month.
Brands that already had a foothold moved on the same timeline. They gained less in raw points, mostly because they had less room to climb.
So a brand's starting tier describes where it begins. Our data suggests it doesn't determine where it ends up, though.
What separated the brands that held a lift from the ones that didn't was volume. Against a control group of 41 brands with no mentions at all, brands with four or more sustained a lift at more than double the rate, a gap that would occur by chance about three times in a thousand.
Four mentions looks like a starting line rather than a threshold. Below it, about one brand in four held its gains. At four and above, more than half did.
You have a brand mass problem, and category-level work has nothing to build on yet. Ranqo's recommendation is to fix that first: a substantial Wikipedia presence, coverage in mainstream outlets, sustained presence on YouTube, etc.
You have a share problem. The two variables are how many of the pages carrying your category's answers include you, and where you sit inside each one. Find where your competitors are being named, and get your name included in those third-party articles, too.
You have a categorization problem, and it's the most fixable. Start with the categories that have no clear owner or a narrow lead between the top two, since those are the contestable ones. Then build the association in other people's content, roughly in this order: "best [category]" roundups, comparison pages against the competitors you want to be grouped with, partner and integration pages, customer stories that name the category, reviews on category-specific sites. The test is whether the model starts naming you when asked the category question.
We opened by saying this research looks contradictory. The truth is that it mostly isn’t. The studies count different things, but once you line the definitions up, they paint a picture about what influences the accuracy of an AI answer, which metrics are worth tracking, how AI recommendations produce real business outcomes, and which sources AI relies on.
Here’s a recap of the numbers that stuck out to us:
Across 23 studies, it all boils down to a few key takeaways. Your own website is somewhere between 1% and 6% of what AI reads about your company. Those pages are still worth getting right because they're what AI draws on when someone asks about your brand specifically. Keeping them current is also half of what decides whether AI states your facts accurately. What they can’t do is get you into the answer when nobody has named you. That happens on the pages AI already cites, and most of those belong to other companies. Which is why the number worth judging this by is how often you appear in those answers, the one measure that holds steady from run to run, and the way to change it is to show up accurately and often in the sources behind them.
Visibility tools only show you the gaps: where you're appearing and where you're not. Noble does something about them, so you show up accurately and more often in AI answers. We find the most relevant sources you’re missing from for your category and key prompts, coordinate with publishers, draft the copy, and make sure it goes live before you pay anything. We work with thousands of publishers to place authentic brand mentions in the third-party sources AI cites, and to correct the ones already out there. You set a budget, and we draw from it only when a mention goes live, so you only pay for success.
GET DISCOVERED ON
