A short brief for campaigns and advocacy organizations: what answer engines actually measure on your pages, why the same gaps lose you human readers, and what to do about it. The research is in the notes.
Ask an answer engine what your candidate thinks about housing. If it hedges, or answers about somebody else, or gives you a sentence that could describe any Democrat in the country, the usual reaction is that the technology failed.
Try the same thing with a person. Hand a stranger your candidate's website and give them nine seconds — roughly what a voter gives it — then ask what your candidate would do about housing. You will often get the same hedge.
That is the argument of this brief. The things that make a page hard for an answer engine to read are, with few exceptions, the things that make it hard for a distracted human to read. Not because the two are alike — they are not — but because both are working under a hard limit on how much they can hold at once, and both are handling prose that never quite says the thing.
"The question is not how do I write for AI. The question is what it means that AI cannot find my candidate's positions on my candidate's own website."
This is worth stating plainly because it changes who owns the work. If AI legibility is a technical problem, it belongs to whoever manages the website. If it is an editorial problem, it belongs to the communications director, and it is the same problem they already have.
Kyanos scores answer-engine responses on four dimensions. They are worth knowing precisely, because each one names a specific way a page fails.
Whether what the engine says about you is true. Wrong positions, wrong biography, endorsements you never received. This measures the information environment around you rather than your own copy — which is why some of it is not yours to fix directly.
Whether the engine can tell who you are. Failure looks like a state senator merged with a U.S. senator of the same name, or accurate facts about an organization that is not yours.
Whether your positions are stated as commitments rather than implied through affiliation or atmosphere. "Supports environmental protection" fails when there is a documented vote you could have named instead.
Whether the claims have evidence underneath them, or are high-sounding language with nothing to check.
Three of those four are entirely within your control. They are written by your staff, on your website, and could be changed this week.
Each of those failures has a matching literature in psychology and communication research. The detail is in the notes; the short version follows.
Naming the subject four different ways makes the reader do your work. A page that says "she," then "the senator," then "the Congresswoman," costs an attentive human almost nothing. But every small resolution adds up, and the capacity it consumes is capacity not spent on your argument.1 Precise, consistent naming also lets a reader file you under something they already understand — a legislator, an environmental group — which is how information gets remembered rather than merely read.2
An implied position cannot be argued with. Persuasion research distinguishes two routes: careful engagement with an argument, and reaction to cues like credibility or affect. The first produces attitudes that last and resist counter-messaging; the second produces attitudes that reverse easily. Careful engagement requires an argument to be present. "Lifelong commitment to working families" gives a reader nothing to think about — there is no claim to accept or reject.3
Prose that is hard to process is judged as less true. This one is unintuitive and well-replicated: material that takes more effort to read is rated less credible, independent of its content.4 Vagueness does not read as harmless. It reads as evasive.
What is missing from your site is chosen by someone else. Framing research holds that influence comes from what is made salient. Answer engines perform their own selection, from whatever material exists. Decline to make your record salient and the selection still happens, from sources you did not write.5
Voters answer with whatever is nearest to hand. A long-standing model of public opinion holds that people do not retrieve settled views; they assemble an answer from whatever considerations are accessible at that moment.6 An AI answer delivered seconds before someone forms an impression is a strong candidate for what is nearest to hand.
Contradiction is worse than silence. People test a story for internal consistency before they test it against the world. A site that presents a candidate as a fighter on one page and a consensus-builder on the next fails that test for a voter and an engine alike.7
| Scoring dimension | What it means on the page | What it costs with people |
|---|---|---|
| Subject clarity | Named consistently, role and place stated | Credibility cannot attach to a referent the reader has to resolve |
| Position explicitness | Specific commitments, named legislation | Nothing to engage with; the reader reacts to cues instead of arguments |
| Factual density | Claims with evidence under them | Reads as assertion; low-effort prose is judged less credible |
| Narrative coherence | The same story across every page | Contradiction, which readers resolve by disengaging |
The column on the right is not a new discipline. It is the job communications directors already have, and have always found hard for ordinary reasons — time, committee approval, the safety of language that cannot be attacked because it does not say anything.
This draws on established research. It reports no experiment of its own. We have not measured whether content edited for answer-engine legibility produces better recall, attitude change, or votes than content that was not. The connections here are inferences from well-supported findings, not results. Treat the argument as a working hypothesis.
Legible prose is not maximally persuasive prose. Emotional resonance, narrative, and authentic voice matter to people in ways no scoring rubric captures, and the persuasion literature is explicit that reaction to cues is real and often decisive — particularly for voters who were never going to study a legislative record.8 Some people vote on how a candidate makes them feel.
The claim is narrower than it may sound: prose that fails on subject clarity, position explicitness and factual density tends to fail human readers on those same dimensions. Fixing that does not produce persuasive writing. It produces writing that is capable of being persuasive.
Nor does all of this work double as editing. Structured data markup, Wikidata entries and machine-readable files are genuinely technical and do nothing for a human reader. The overlap argued here covers the editorial dimensions, not the plumbing.
The practical value here is not that scoring tells you how to write. It is that scoring makes a familiar failure visible and countable, on a specific page, in a way an internal review rarely does. Nobody circulates a memo saying the issues pages are vague. A score does.
Three things follow, in order of how quickly they can be done.
Read your own issues pages against one question: what exactly does this say, and about whom? Most of what fails is not false. It is diffuse — true sentences that commit to nothing. That review costs a morning and needs no vendor.
Name the subject the same way everywhere, including role and jurisdiction, on every page. This is the cheapest fix available and it addresses the failure most likely to get your candidate confused with someone else.
Replace one implied position per page with a specific one — the bill, the vote, the number. If a claim cannot be made specific, that is worth knowing internally, and it is a different conversation than an editing one.
None of this is new advice. It is the advice every communications shop has already been given and mostly not taken, because the cost of vagueness was invisible. What has changed is that a machine now reads the page first, answers on your behalf, and does it identically every time — which makes the cost measurable for the first time.