Does an author bio affect AI citation?

Research question: does a byline change citation?

The question: if you add a named author bio with visible credentials to a page that currently has none, and change nothing else, does that page get cited more often by AI search engines?

Why it is unanswered: E-E-A-T's evidentiary basis is Google's rater guidelines for human quality evaluation, not a disclosed AI-citation study. Google's own AI optimization guidance and its documentation on AI features describe content quality in general terms and name no author signal specifically. The extension from "helps human raters judge quality" to "increases AI citation" is plausible and entirely untested, which is why it is graded Hypothesis on the tactic evidence scoreboard. This study is designed to move it off that grade in either direction.

What evidence exists today?

What exists is correlational and vendor-published. Zyppy has ranked candidate citation factors, Ziptie has argued that original research wins citations and described how ChatGPT appears to choose sources, and Discovered Labs has compared citation patterns across platforms. None isolates authorship, and none discloses a corpus — the same provenance problem running through every widely-quoted figure in this field and the reason the measurement standard exists. The academic GEO paper that named the field tests content-level interventions and says nothing about authorship at all.

Where did E-E-A-T actually come from?

The lineage matters for interpreting whatever this study finds. It began as E-A-T in Google's publicly released Search Quality Rater Guidelines — a document used to train human evaluators who judge search result quality, not an automated ranking signal. It is an input to checking whether Google's ranking produces good results. A fourth "Experience" dimension was added later. Turning that into an automatable AI-citation signal is an extension the industry made by analogy, not a claim any operator has validated.

Four channels a credential could travel through

"Does E-E-A-T help?" is too coarse to test. A credential can reach a retrieval system by at least four routes, each needing a different experiment.

Most public advice blends all four into one recommendation, which is exactly why that advice is unfalsifiable: a site that adds bylines, builds author profiles elsewhere and grows its domain reputation over a year cannot say which did anything. Hypothesis Our prior is that off-site corroboration is the strongest channel, because it is the one with independent evidence behind it — but that is reasoning, not data, and it is the hardest of the four to randomise.

Method: sample, control and randomisation

Sample. 300 pages across recruited participant sites, all currently published with no visible author byline, and all eligible for a real named author to stand behind them. Randomisation is stratified by topic category so treatment pages cannot skew toward topics with higher baseline citation.

Control. 150 pages left anonymous and otherwise untouched for the same window. The control arm is what makes a mid-study engine change survivable: if retrieval shifts under everyone, both arms move together and the comparison holds.

Variables. The manipulated variable is the presence of the standard bio template. Measured outcome is citation rate against a query set frozen at registration, collected per engine and never blended — the surfaces already disagree, as the AI Mode versus AI Overviews source delta shows. Crawler activity is logged alongside, so no page is measured before it has been refetched. The design is the same shape as the llms.txt test.

What counts as a qualifying bio?

Every treatment page receives the same template, to keep the treatment consistent and prevent post-hoc redefinition: a full name, a one-line stated credential or role relevant to the page's topic, and a single sentence of stated experience. No external verification links, no structured markup change, and no other edit to the body content. Deliberately excluded are any change to structured markup or to publication dates, both registered as their own experiments. Holding the template constant across all 150 treatment pages is what lets a measured difference be attributed to the presence of a bio rather than to how elaborate any individual bio was.

Hypothesis: the single registered prediction

H1: Adding a named author bio with visible credentials to a previously anonymous page causally increases its AI citation rate within one collection cycle, holding the underlying content identical. One primary outcome, named before collection; every other slice — by engine, topic or window — is labelled exploratory when reported.

What would a result look like?

Here is the measurement, with invented, hypothetical numbers only. Say a panel page on "how to choose a business insurance policy" previously published with no visible author name, sits at a 4% citation rate across the tracked query set. It is randomized into the treatment group and receives the standard bio template. No other change.

One collection cycle later, its citation rate measures 9%. Its matched control page, left anonymous, moves from 4% to 5% over the same window, within normal noise. If that pattern held consistently across the full 300-page panel it would support H1; if bio and non-bio pages moved indistinguishably, that would support the null instead. Both outcomes publish.

Either way, a mechanism is worth naming in advance. A positive effect would suggest visible expertise signals help a retrieval system separate a trustworthy source from a competing anonymous page. A null would be consistent with current-generation retrieval not parsing bio content as a distinct signal at all. That matches the live-fetch finding that systems often extract only a page's primary visible content — the same main-content bias Onely describes from the other direction in what makes content LLM-friendly. It also sits next to the open question of whether those systems execute JavaScript at all.

Control: confounds this design has to defeat

Randomisation handles a great deal, not everything. These threats are stated in advance so a reader can check whether we managed them.

Concurrent site changes. Panel sites keep publishing and redesigning during the window. Randomising across sites reduces this, but a large change on one site can still move the numbers.

Engine changes mid-study. Retrieval systems update without notice. The control arm is what protects the finding, which is exactly why the study has one.

Topic and vertical imbalance. Citation behaviour differs enough between local and commerce queries that an unbalanced panel would produce a topic effect wearing a bio's clothes. Stratified randomisation is the fix.

Recrawl timing. A page not recrawled since the bio was added cannot show an effect; measuring before recrawl measures latency instead. That is why the crawl-to-citation latency work matters here and why crawler activity is logged per page. For scale, Ahrefs found conventional ranking itself takes months; there is no reason to assume citation moves faster.

Query set drift. If the tracked set changes composition between cycles, rates move for reasons unrelated to any page. The set is frozen at registration.

Analysis: measurement traps in a citation RCT

Beyond confounds are the ways to run this correctly and still report it wrongly. Each of these is a rule the analysis commits to now.

Persistence is a separate question this design does not answer: a change measured once says nothing about how long it lasts, which is the subject of the citation half-life study. A citation result is also not a traffic result, so nothing here converts directly into a business case.

Why a result would not transfer between engines

Even a clean positive would describe one pipeline at one moment. Retrieval systems differ in what part of a page they read: some may extract only the main content block and discard the surrounding furniture, which is exactly where a bio usually sits, while others process the full document. That single architectural difference could produce opposite results on the same page.

They also differ in how much they lean on an underlying index versus fetching live — Seer found 87% of SearchGPT citations matching Bing's top results, while Ahrefs reports far lower overlap with Google's. An engine inheriting its candidate set from a classic index could therefore show an E-E-A-T effect arriving second-hand from classic search rather than from AI retrieval at all. An engine fetching live with its own retrieval fleet could not.

Per-engine profiles already diverge. ChatGPT, Perplexity, Claude and Gemini each favour different source types, as Discovered Labs, Leapd and Profound all report, and the concordance work exists to measure that spread. Open question Whether a bio effect found on one engine replicates on another can only be answered by measuring each separately, which is why no blended headline number will be published.

Interpretation: what each outcome would license

One further misreading is worth pre-empting: "a positive result means add bios everywhere." Only where a real person genuinely stands behind the page. A byline on content nobody authored is fabrication, and it fails for reasons that have nothing to do with citation rates — most sharply in YMYL categories, where authorship carries obligations independent of any engine.

The ethics of a bio experiment

An experiment that adds author credentials carries an obligation most SEO tests do not. If the credentials are not true, the study manufactures exactly the misleading signal this site exists to argue against — the standard set out in how this publication works. So the panel rule is simple: every bio names a real person who genuinely stands behind the content, with a credential they actually hold. No invented experts, no borrowed titles. A page whose site cannot supply a real author is not eligible for the treatment arm.

That constraint costs the study something — it narrows the eligible pool and slows recruitment — and it is the correct trade, because a result obtained by fabricating credentials would be worthless even with perfect statistics. A second obligation runs to readers of the panel pages. A bio added for a study is still a bio a reader will rely on, so it has to be accurate on its own terms and stay accurate after the experiment ends.

Null results we would publish

The pre-commitment is the point of registering publicly. A null here is not a failed study; it is the more likely outcome and arguably the more useful one. We would publish treated and control pages moving indistinguishably, with the confidence interval alongside so a reader can see the range of effects the data cannot rule out. We would publish a result positive on one engine and null on the others, without averaging them into a headline. We would publish a result too underpowered to conclude anything, with the power calculation shown.

And we would publish a failure to complete: insufficient recruitment, excessive attrition, or a mid-study engine change severe enough to invalidate the window. Registering a study and reporting that it could not be run is more honest than quietly dropping it. Each outcome lands in the null results registry.

Reproduction: running a smaller version yourself

You do not have to wait for a 300-page panel to learn something, though a single-site version is weaker. Scripting it is covered in the Claude Code for SEO guide.

Be honest about what that delivers. With a small sample on one site, only a large effect clears the noise: a flat result tells you very little, and a striking one tells you it is worth testing properly. The bots to watch for are listed in the AI bot user-agent registry, which also covers how often sites block them by accident.

Who this applies to, and who it does not

Schedule and raw data

Recruitment runs alongside the schema and freshness RCT panels, targeting Q1 2027 for the first collection cycle. Published with the result: the randomisation assignment list, the frozen query set, the per-page citation checks including every negative, and the crawl log confirming refetch. The bio template itself is already fixed above, so the treatment can be inspected before anyone sees an outcome. Terms used here are defined in the AI search glossary. Whether this counts as GEO, AEO or LLMO work changes nothing about the evidence.

Limitations

• Sample: recruited volunteers, 300 pages. Participant sites self-select, which skews toward publishers already engaged with AI-search questions, and eligibility requires a real author willing to be named — narrowing the pool further, plausibly toward better-resourced editorial operations.
• One implementation of "author bio" is tested. A named byline with a stated credential, in one fixed template. Structured Person markup, external credential verification and richer bios are untested, and credential strength is held constant rather than varied, so nothing here speaks to whether a more impressive credential changes the effect size.
• Geography and language are uncontrolled. The panel will not be balanced across regions or languages, and author-signal conventions differ between them, so the result describes the panel's markets rather than the web's.
• Engine coverage is limited to surfaces with a checkable citation list. Agent-facing routes — an MCP server, a WebMCP endpoint or an agentic browser — may never see a rendered bio at all and are outside the design.
• Query selection bounds the outcome. Citation rate is measured against a query set frozen at registration; a page could gain citations for queries outside it and register as unchanged.
• Measurement depends on recrawl, which is not controlled. Pages fetched infrequently may show no effect simply because the treatment was never seen, and attrition from that cause could differ between arms.
• Confounders remain: freshness and off-site footprint. A bio added during any wider site refresh is impossible to separate from recency effects, and an author's off-site presence — the channel we suspect is strongest — cannot be randomised or held constant.
• Reproducibility is bounded by panel privacy. Sites that ask not to be named appear in the raw data by identifier only, so an independent replication can check the analysis but not re-observe the same pages.
• Generalisation is single-cycle and time-bounded. The design cannot say whether any effect decays after the first recrawl, or whether it holds once engines change, and it can say nothing about whether credential relevance to the topic matters as distinct from credential presence.