Study #1 · October 2026

We tested whether PR moves AI visibility. Most of the signal is shared with brand size.

We compared how many leading news and trade outlets link to 120 brands with how often ChatGPT, Google AI Overviews and Google AI Mode mention them. The two move together, but most of that link is shared with brand size, and a cross-section cannot say how much of that size coverage itself helped build. New coverage did not predict AI share gains, and news sites are about 8% of what the engines cite (a floor).

Green Flag Digital · 120 brands · 3 AI engines · pre-registered, every number re-run by an independent verifier

Key findings

  1. Leading-outlet links and AI mentions move together. Pooled within-vertical rho is 0.53 (95% CI 0.37 to 0.65), and it is the same in ChatGPT (0.53), AI Overviews (0.52) and AI Mode (0.53).
  2. Most of that is shared with brand size. After accounting for brand size and brand search demand, outlet links add little that we could detect: partial rho 0.18 (95% CI -0.05 to 0.37), or 0.39 if only brand demand is controlled. The study cannot rule out a small effect, and because coverage can also help build size, 0.18 is a lower bound.
  3. New coverage did not predict AI share gains. Brands that gained more new outlet links in the last six months did not gain more AI share (partial rho 0.04, 95% CI -0.15 to 0.24).
  4. Editorial counts track AI mentions more closely than wire counts. Rho 0.53 for editorial outlets versus 0.21 for press-release wires, a difference of 0.33 (95% CI 0.09 to 0.55). Much of that gap reflects the fact that bigger brands get more editorial coverage.
  5. News is a small slice of what AI engines cite. News and magazine sites are about 8% of cited sources on ChatGPT, AI Overviews and AI Mode (a floor). ChatGPT leans on the sampled brands' own sites (34%); Google's AI Overviews and AI Mode lean on Reddit, YouTube and other user content (25-30%).

All findings are associations. This is a cross-section of 120 brands, and it cannot show that coverage causes AI visibility, or the reverse.

Why we ran this

Every PR team is being asked the same question right now: does coverage get my client into AI answers? Most of the public answers are confident and unsupported. Ours was a guess too, so we wrote down the tests before pulling the data, ran them, and agreed to report whatever came out.

Our core hypothesis, that coverage predicts AI visibility beyond brand size, failed. So did the timing test and our guess about what AI engines cite. We are publishing all of it. A study that only reports the flattering result is a sales page.

“We wrote down our hypothesis before we looked at the data, and it failed. We are publishing it anyway, because an agency that only reports its wins is asking you to trust its losses to luck.”
Joe Robison, CEO, Green Flag Digital

What we measured (plain English)

The brands. 120 US-facing brands, 30 each in cybersecurity, travel, legal (legal tech and consumer legal) and B2B SaaS. Each vertical runs from category leaders to mid-market names. The sample was frozen and hashed before we pulled any AI data.

AI visibility. Ahrefs Brand Radar counts of how often each brand is mentioned in answers from ChatGPT, Google AI Overviews and Google AI Mode. We summed the three. Perplexity, Gemini and Copilot are not licensed on our plan, so they are not in the study.

Coverage. The number of distinct outlets, from a fixed list of 68 national, business, tech and trade publications, that carry a live link to the brand. This counts links from leading outlets. It is not verified earned coverage, and it says nothing about volume, sentiment or recency. Contributor networks, syndicators, wires and review sites are excluded from the main measure.

Controls. Referring domains outside those outlets, organic traffic, paid traffic and US search volume for the brand name.

How we compare. Rank correlations within each vertical, so a travel brand is only compared with other travel brands. A rho near 0 means no relationship; near 1 means brands rank the same on both measures.

Finding 1: Leading-outlet links and AI mentions move together

Brands linked from more leading news and trade outlets tend to be mentioned more by AI engines. The pooled within-vertical rho is 0.53 (95% CI 0.37 to 0.65). ChatGPT, AI Overviews and AI Mode each land within a point of that figure.

By vertical, the point estimates are 0.55 for cybersecurity, 0.73 for travel, 0.54 for B2B SaaS and 0.32 for legal (95% CI -0.08 to 0.64). With 30 brands each, those vertical numbers are descriptive. We do not rank verticals against each other.

Scatter of outlet-link rank against AI-mention rank, by vertical
Scatter of outlet-link rank against AI-mention rank, by vertical
Forest plot of rho by engine and vertical
Forest plot of rho by engine and vertical

What it means for marketers: If you want to know which brands an AI engine currently talks about, look at who leading outlets link to. Do not read that as a recipe, because the next finding explains most of it.

Finding 2: Most of the link is shared with brand size

Big, established brands get covered and get mentioned. Both can follow from being big, though coverage may also be one of the ways brands get big (see Limits). When we controlled for size, traffic, ad spend and how often people search the brand name, the pooled partial rho dropped from 0.53 to 0.18 (95% CI -0.05 to 0.37), which we cannot distinguish from zero. Controlling for brand search alone brought it to 0.39 (95% CI 0.20 to 0.53). The pre-registered size-controlled test did not pass our multiple-testing correction.

Even the 0.18 is fragile. Brands with ambiguous names, such as "Going," get inflated AI counts from the common word. Remove the 18 ambiguous brands and the partial rho is 0.01 (95% CI -0.22 to 0.24).

We also ran a placebo: trade outlets from the wrong vertical, such as legal trade press linking to a cyber vendor. A placebo exposure shows a similar raw association (0.45, 95% CI 0.28 to 0.59) that also disappears after controls (0.05, 95% CI -0.17 to 0.27). That is what you would expect if the raw number mostly measures brand size.

Power matters here. A true partial effect below about 0.25 is compatible with our data. "Not detected" is not "no effect."

Correlation shrinks as size controls are added, with placebo
Correlation shrinks as size controls are added, with placebo

What it means for marketers: Do not expect PR to stand in for brand size. In this sample (no true small challengers), the brands that win both press and AI answers tend to be the ones that were already winning.

Finding 3: New coverage did not predict AI share gains

If coverage moved AI visibility, brands that picked up more new outlet links over six months should have gained AI share. We could not detect that they did (partial rho 0.04, 95% CI −0.15 to 0.24).

There is a weak hint in the other direction: brands already visible in AI in April went on to pick up more new outlet links (rho 0.19, 95% CI 0.01 to 0.36; unadjusted, exploratory). Not detected is not the same as no effect, and six months is a short window. The link date we use is when Ahrefs first crawled the link, not the publication date, and the engines' prompt panels grew unevenly over the window, which adds noise.

What it means for marketers: Do not expect a single quarter of placements to show up as an AI share gain. Measure it, but set expectations low.

Finding 4: Editorial counts track AI mentions more closely than wire counts

Counting distinct outlets, editorial links correlate with AI mentions at rho 0.53 (95% CI 0.37 to 0.65). The three press-release wires correlate at 0.21 (95% CI 0.01 to 0.38). The difference is 0.33 (95% CI 0.09 to 0.55), and it passed our multiple-testing correction (adjusted p 0.016) and every pre-registered sensitivity run.

The important caveat: much of that gap reflects the fact that bigger brands get more editorial coverage. Editorial-outlet count tracks brand size closely; wire pickup does not. When our verifier added the same size controls, the gap was no longer distinguishable from zero (0.12, 95% CI -0.12 to 0.36). Yahoo also syndicates wire copy, which blurs the contrast. Using wire links instead of wire outlets (the pre-registered secondary unit), the gap narrows to 0.17 (95% CI −0.00 to 0.34). The wire count is also near its ceiling (96 of 120 brands have 2 or 3 of the 3 wires), which flattens its correlation.

Editorial versus wire-outlet correlation with AI mentions
Editorial versus wire-outlet correlation with AI mentions

What it means for marketers: Wire pickup on its own is the weakest coverage signal in this data. Editorial coverage is the stronger signal, mostly because it comes with being a brand editors care about.

Finding 5: News is a small slice of what AI engines cite

We guessed that news and trade sites would be the biggest outside source AI engines cite. They are not, on any engine. Across answers that mention 100 brands with unambiguous names, news and magazine sites are about 8% of cited sources on ChatGPT (7.9%), AI Overviews (7.5%) and AI Mode (7.8%). By vertical (exploratory, descriptive only), the earned share runs 11-15% in cybersecurity, 6-10% in travel and 4.5-5.6% in legal and B2B SaaS across the three engines.

The engines differ in where they go instead. ChatGPT leans on the sampled brands' own sites (33.6%). Google's AI Overviews and AI Mode lean on Reddit, YouTube and other user content (30.2% and 25.0%). Review and directory sites are 7-9% everywhere. The biggest bucket, about 35-39%, is a mix of other companies' sites, travel tools and affiliate sites.

Two notes. The 8% is a floor: when our verifier blind-coded a sample of domains (agreement kappa 0.86), about a fifth of the "other" bucket turned out to be review or news sites, so the true earned share is plausibly 8-10%. And this was an exploratory test.

On the earlier question of engines: the link between outlet links and mentions looked the same in ChatGPT and AI Overviews (difference 0.008, 95% CI -0.05 to 0.07; exploratory). Engine mention rankings are highly correlated with each other, so that is not surprising.

Share of AI citations by source type, by engine
Share of AI citations by source type, by engine

What it means for marketers: AEO is not just a PR problem. On ChatGPT, your own pages carry the most weight. On Google's AI surfaces, community and video content does. News and trade outlets are one cited source of several.

“You cannot buy your way into AI answers with volume. Big brands get cited because they are big, and coverage alone did not close that gap in our data. Specific coverage, in the outlets the engines cite for your buyers' questions, measured before you start, is the only version of this we would sell.”
Joe Robison, CEO, Green Flag Digital

What this means for your PR and AEO strategy

  1. Treat PR as brand-building, not an AI hack. The data supports "the brands AI recommends are the brands editors cover." It does not support "PR causes AI visibility."
  2. Don't count on wire pickup alone. It was the weakest coverage signal we measured. Aim for coverage editors choose to write.
  3. Work the sources the engines actually cite. Your own pages (ChatGPT's biggest source), Reddit and YouTube (Google AI's biggest sources), and review sites. Press is one of several inputs.
  4. Judge placements over quarters, not weeks. Six months of new coverage did not predict AI share gains in this data.
  5. Fix your own entity first. 18 of our 120 brands have ambiguous names, and several (such as Going and Bill) had clearly inflated AI counts. If your brand name is a common word or shares a name with another entity, check what the engines actually say about you before you read any dashboard.
“We changed how we sell. Every program now starts with a day-0 scoreboard and a margin of error, and if the panel has not moved 2 points by month 3, either side can walk.”
Joe Robison, CEO, Green Flag Digital

Methodology box

Sample. 120 brands, 30 per vertical (cybersecurity, travel, legal, B2B SaaS), spanning leaders to mid-market by referring domains. The smallest brand has 2,090 referring domains, so true small challengers are absent. The sample was frozen and hashed before any AI pull.

Engines. ChatGPT, Google AI Overviews, Google AI Mode, via Ahrefs Brand Radar, US only. Perplexity, Gemini and Copilot are not covered. Primary outcome is mentions summed across the three.

Exposure. Distinct outlets with a live link to the brand, from a strict pre-listed set of 68 national, business, tech and trade outlets (the full 84-outlet list was a sensitivity run). Median 22 outlets per brand (range 3 to 51). Links can be sponsored, contributor-written, syndicated or self-published, so this is "linked from leading outlets," not "earned coverage."

Tests. Within-vertical rank correlations with bootstrap 95% confidence intervals. Size-controlled analysis with a negative-binomial model and a rank-based partial correlation, plus log-OLS as a check. The four confirmatory tests (pooled correlation, size-controlled effect, new-coverage test, editorial versus wire) were Holm-adjusted. The pooled correlation is descriptive and cannot be a headline on its own. Engine splits, vertical splits and the engine comparison are exploratory. Reliability check: the two Ahrefs endpoints rank brands almost identically (Spearman 0.99 for ChatGPT, 1.00 for AI Overviews).

Pre-registration. The rules were written before the main data pull and are in METHOD.md. Changes after that are disclosed:

  • Before the analysis run, we added a sensitivity run that also drops brands whose name variants match unrelated things (for example, "Checkpoint" matching airport security).
  • An amendment adopted before the main pull tightened the exposure list to 68 outlets, set the primary outcome to the 3-engine sum, and fixed the control set and the Holm family.
  • During the first analysis run, the negative-binomial solver failed on heavy-tailed counts and was replaced with maximum likelihood. We then used a 2,000-resample bootstrap for that model's p-value, after seeing its first result (disclosed). An independent verifier found and we fixed a bug that dropped non-random resamples. The size-controlled test fails the multiple-testing correction either way (adjusted p 0.066 for the model, 0.25 for the rank-based version), and the model's estimate is driven by ambiguous brand names.
  • The first cited-sources pull was contaminated by ambiguous brand names (Home Depot and Walmart cited for "Bill" and "Going"), so we re-pulled it on 100 unambiguous brands and weighted the four verticals equally.

Limits.

  • Correlation, not causation. Reverse causality cannot be excluded.
  • Ahrefs supplies both the links and the mentions, so shared crawl-coverage bias is possible.
  • Power: the minimum detectable rho is about 0.26 pooled and about 0.50 per vertical.
  • Range restriction (no true small challengers) biases correlations toward zero.
  • Over-control. Referring domains, organic traffic and brand-name search volume are treated as measures of size, but years of coverage can also build them. If coverage is one of the ways brands get big, these controls absorb part of its effect, so 0.18 is a lower bound and the raw 0.53 an upper bound. A cross-section cannot separate 'size causes both' from 'coverage helped build the size'.
  • 18 of 120 brand names are ambiguous, and their AI counts may be inflated. Headlines are checked with them removed.
  • The ChatGPT prompt panel grew unevenly between April and October 2026.
  • 3 of 6 major AI engines were tested.
  • Cited-source shares use the top 200 domains per vertical and engine; the earned-media share is a floor.
  • The exposure is a count of links from a curated list. It is not coverage volume, sentiment or recency.

Data and code: the method and charts are published here. The pre-registration, the reproduction script and the derived brand-level tables are available on request; raw Ahrefs API responses are not redistributed (vendor data). Reproduction script: `analysis/run_analysis.py`. An independent verifier re-ran every number from the raw data, checked all 673 raw API files against the dataset, and blind double-coded the cited-source categories.

About GFD

Green Flag Digital is a data-driven PR agency for B2B software companies. It has done SEO and digital PR for clients since 2015, and runs every program on GFD ONE, its own system for measuring and moving a brand's share of AI answers. https://one.greenflagdigital.com/services/digital-pr

Your numbers

See where your brand stands in AI answers.

Tell us where to look. We come back with your share of AI answers, your top competitor's, and the pages we would take first, then walk you through it in 20 minutes. Free.

Get your free Quick Look