Launch pricing · 25% off every listingBrowse domains
RevisedRevised
Browse domains
ai-searchdomain-authoritybacklinksseo

Introducing Agent Citability: A Domain Metric for the AI Search Era

Domain Rating can be bought for $15 on Fiverr and it only ever goes up. Agent Citability weights links by who is doing the linking, and it can fall.

August 22, 2026

Every domain we list gets the same question: what's the DA?

We've answered it thousands of times, and we've come to think it's the wrong question. Not because the number is hard to find, but because of what has happened to it. So we're building a different one.

We're calling it Agent Citability.

Domain Rating is a number you can buy

Moz and Ahrefs built DA and DR to approximate PageRank, back when that was a reasonable ambition. Then the web did what it does to any public, stable scoring function. It optimized against it: directory entries, comment spam, private blog networks, and eventually a retail market in the score itself.

Max Roslyakov of Xamsor priced that market. Across seven experiments covering more than 12,000 domains, he paid Fiverr contractors $275 in total to inflate authority metrics on freshly registered domains. Individual domains cost $15 to $85. DR on those test domains landed between 44 and 59.

That is premium guest-post territory, on domains that were days old and had earned nothing.

The second finding should end the conversation. Roslyakov documented hundreds of cases after Google's March 2024 update where a site was penalized into near-zero traffic while its DR and DA held steady or climbed.

The search engine the metric exists to approximate has decided the site is worthless, and the metric gets more enthusiastic. That isn't an implementation bug. These scores are computed from link-graph shape, and a link graph has no way to represent trust being withdrawn. They go up. They have no mechanism for going down.

We've written before on what DA does and doesn't measure. It's still fine for rough competitive comparison. As a statement about whether anyone should trust a domain, it stopped working a while ago.

Then the index changed underneath us

While the metrics decayed, the surface they proxied for stopped being the only one. Answer engines now sit between a lot of people and the websites they'd otherwise visit, and what those engines cite is not what classical authority ranks.

Analyze.ai studied 22,295 AI answers producing 115,843 citation events across ChatGPT, Perplexity and Google AI Mode. It found 7,058 unique cited domains, with the top ten accounting for only 11 to 13% of citations on each engine. Around 88% of citations sit outside the top ten on all three. (That study runs on business-software prompts, so treat it as evidence about commercial queries rather than all of search.)

A Washington University measurement study points the same way. Xu, Iqbal and Montgomery measured 55,393 trending queries over 40 days and found that nearly 30% of the domains cited in an AI Overview don't appear in the first-page results shown beside it, and that cited domains are more credible than those co-displayed results. Whatever selects AI citations is not the ranking algorithm.

That long tail is good news if you own a specific, unglamorous domain about one subject. Niche sites get cited. Being enormous is not the entry requirement.

What the link graph still tells you, and what it doesn't

Here's where we have to be careful, because the research genuinely disagrees with itself and we'd rather say so than cherry-pick.

Ahrefs tested 75,000 brands for what predicts showing up in Google AI Overviews. Their own metric placed mid-table:

SignalCorrelation
Branded web mentions0.664
Branded anchors0.527
Branded search volume0.392
Domain Rating0.326
Referring domains0.295
Backlinks0.218

Ahrefs are careful that these are moderate-to-weak correlations, and they filtered to domains above DR 40, which compresses DR's range and works against it.

Pointing the other way: SE Ranking analyzed 129,000 domains and 216,524 pages and found referring-domain count the strongest single predictor of ChatGPT citations, with sites under 2,500 referring domains averaging 1.6 to 1.8 citations against 8.4 for those above 350,000. Though a domain with 350,000 referring domains is a household name by definition, and that study controls for nothing else, which is rather the point.

So we're not going to tell you links don't matter. Look again at where branded anchors land: 0.527, against 0.218 for raw backlinks. Branded anchors are links, since anchor text is a property of a backlink. The difference is that they carry a brand name, and every brand-carrying signal in that table outranks every pure link-count signal. What the link graph predicts is mostly this is a big, old, well-known brand, which is exactly what DA and DR already price, and exactly what a decade of link buying learned to counterfeit for $15.

The differentiator isn't whether you count links. It's which ones.

What Agent Citability measures

Agent Citability is a 0 to 100 score for how deeply a domain is woven into the trusted layer of the web.

It runs on a curated, tiered list of over 1,000 sources that AI answer engines demonstrably cite. A citation from Wikipedia, a link from an academic paper, a reference from a .edu page, or a link from a newsroom of the NYT, CNN or Washington Post class carries real weight. A user-submitted directory listing carries nothing. Not a little. Nothing.

One caveat on that, stated now rather than left for you to find: several major newsrooms restrict automated crawling, so their pages emit almost no links into the current release. Newsroom weight is real in the model and close to inert in today's preview, which is why the live signal leans on reference, government and academic sources instead. More news seeds won't fix that; broader crawl coverage will. The methodology page carries the detail.

That tiering is the whole design. Scoring domains on raw citation counts would be farmable on day one, because the most-cited domains in Google AI Overviews are YouTube, Reddit, Facebook, Instagram and Quora, and anyone can publish a link on those in about a minute. What matters is whether the citing domain chose to publish the link.

The limit is in the name. Agent Citability measures citability, not citations. It does not predict that a domain will be cited, and given how dispersed and how engine-specific the citation data is, anyone promising that is selling something. It measures provenance, which is an input rather than an outcome.

How it's computed

Agent Citability uses personalized PageRank over our domain-level web-graph index, which covers more than 100 million domains and billions of links between them. Rather than treating every link as equal, trust propagates outward from the seed set, following the TrustRank method published by Gyöngyi, Garcia-Molina and Pedersen in 2004.

Scores are recomputed each time the graph is rebuilt, and calibrated to a percentile within that release.

The methodology gets published too. The exact tier weights don't, and it's worth saying why plainly. Everything in the first half of this post is what happens when a scoring function is public, stable, and cheap to target. Publishing weights hands a spec sheet to the people selling DR 50 for $15. Publishing the method lets you audit what we claim to measure and argue with the tiering.

What it doesn't let you do is recompute our number. Holding the weights back means you can check the ingredients and the machinery but can't reproduce the output, and we'd rather say that than imply an openness we don't have.

The full method, and the seed-set criteria as they firm up, live on the Agent Citability methodology page.

The study that argues against us

In July 2026 Surfer analyzed about 5 million citation sources across 20,000 prompts and four platforms, and found link authority barely correlates with getting cited: Spearman -0.065 for PageRank, +0.010 for harmonic centrality, -0.051 for their Domain Score. Their words: "close enough to zero to be noise."

We'd rather raise that ourselves than have you find it. If link authority is noise against AI citation, a link-based metric looks like a bad idea. What they measured is citation frequency among pages already cited, which is not the same quantity as citability, and we'd be fooling ourselves to treat that distinction as a defence before testing it.

Two things make us think it's worth testing anyway. Their population is URLs already cited at least once, so correlating against citation frequency conditions on the outcome and attenuates whatever signal exists. And they tested global PageRank, not trust seeded from a curated set, which is a different computation over the same graph.

Both of those are hypotheses we're testing, not results we have. If seeded trust turns out to be noise too, that's a finding we'll publish.

What's actually new here, and what isn't

Not the idea of a domain-level trust score. Majestic have run Trust Flow, computed as TrustRank from a curated seed set on a 0 to 100 scale, since 2011. Domain-level PageRank and harmonic centrality have been published from web-scale graph data for years, and DomCop has served an API over that data since 2018.

What's new is the seed set and the disclosure. The seed is derived from what AI answer engines actually cite rather than from general reputability, and we publish the method. Nobody in the AI-visibility category does the latter; Semrush call theirs proprietary outright.

That category is also solving a different problem. Every funded platform we checked scores a brand against a configured prompt corpus. Ask one about an arbitrary domain that isn't a tracked brand and there's no number to give you.

The scores will go down

Agent Citability ships as versioned releases (Agent Citability 2026-Q3, then the next), each recalibrated, each with a published changelog.

Scores will fall between releases. A domain drops if the trusted tier stops referencing it, if recalibration corrects something we got wrong, or if the percentile shifts under it. We're not promising the scale stays linear or that a score means the same thing across versions, which is the mistake Moz made re-basing DA, and the changelog exists so you can see what moved.

That's the feature. The specific failure of DA and DR is that nothing in their construction can register withdrawn trust. A number that only ratchets upward is a marketing number, and if we shipped one we'd have rebuilt the thing we just spent five paragraphs criticizing, in a nicer typeface.

It cuts against us. We sell domains, and a score that can drop after purchase is not commercially convenient. We'd rather carry that than publish another number nobody should believe.

What's live today

A preview is live as of today. Listings across the marketplace now carry an Agent Citability score wherever the domain appears in the current release, which is roughly three in five listings today, and that share grows as we fold in the earlier releases we already hold. The free domain report and the Agent Citability checker both show a score for any domain with trusted-source links behind it, whether or not you've bought anything from us. Arguing the industry's metrics are broken and then charging for ours would be a poor look.

Treat these as v0 preview numbers and nothing firmer. The seed set is still being tuned, the scores haven't been through a full calibration pass, and they will move. The first proper release, with version notes and a changelog, is still ahead of us, and we're not putting a date on it.

If you want to know whether this holds up, watch the methodology page rather than this post.

Sources