Contradiction Debt: Fixing Conflicting Facts About You Across the Web

Technical debt slows your codebase. Contradiction debt slows your entity. The difference is that you can see the first one, while the second compounds silently in directories you forgot existed.

What is contradiction debt?

Contradiction debt is the accumulated set of conflicting facts about you across the web: two different job titles, three bio variants, a company you left in 2022 still listed as current, a speaker page with the wrong credential, a name spelled two ways. Each conflict was cheap to create, usually by simply not updating something, and each one levies a recurring tax: it lowers the confidence with which a machine can assert who you are. When an engine assembles an answer about a person, it is fundamentally reconciling sources. Sources that agree produce a crisp, confident answer. Sources that disagree produce hedges, blends of old and new facts, or silence.

Silence is the expensive part. A recommendation engine choosing between two comparable experts will favor the one whose record reconciles cleanly, for the same reason a cautious journalist quotes the source whose story checks out. You do not get an error message when this happens. You just do not get named.

The hedge tax: how contradiction debt shows up in a generated answer

Contradiction debt is invisible until you know what to look for, and what to look for is hedging. When a model retrieves several passages that disagree about a fact, it cannot silently pick a winner the way a database join would. It generates language, and generating language across disagreeing evidence produces a specific, recognizable pattern: qualifiers, blended claims, and vague attribution.

What a clean answer looks like

When every source agrees, the model states the fact plainly: "Maria Costa is the head of partnerships at her firm, focused on mid-market retail accounts." No qualifiers, no hedging, a single confident sentence built from converging evidence.

What a contradiction-debt answer looks like

When sources disagree, the same question produces something like: "Maria Costa has been associated with partnerships roles, and some sources describe her focus as retail while others mention broader account work." Notice the shift: "has been associated with" instead of "is," "some sources" instead of a name, and a blended description that satisfies neither version. That sentence is not a style choice. It is the visible output of a model averaging across passages that would not reconcile.

The tax compounds in a second way too. A hedge is a weaker sentence, and weaker sentences are less likely to be the one a model chooses to foreground when a user asks a comparative question like "who should I talk to about X." Between a confident, cleanly sourced answer about a competitor and a hedged, blended answer about you, the confident one tends to win the recommendation, not because the competitor is better, but because their record reconciles and yours does not.

Where do conflicting facts about you come from?

Audit enough professionals and the same seven leak sources appear:

• Stale profiles: the conference speaker page from 2021, the association directory, the alumni listing, all still asserting the person you used to be.
• Old press: accurate when written, wrong now. Journalistic archives never update themselves.
• Ghostwritten bios: every event organizer who "tightened up" your bio shipped a slightly different version of your title and claims.
• Abandoned accounts: the dormant Twitter alt, the old Medium, the agency-era portfolio site still live on a forgotten subdomain.
• Name variants: Jonathan on LinkedIn, Jon in bylines, J.R. on the paper you published. Humans reconcile these instantly; resolvers only sometimes do.
• Scraper sites: data aggregators that copied an old profile and now republish it with total confidence and zero maintenance.
• Your own site: the About page says one thing, the footer bio another, the media kit PDF a third. The canonical source contradicting itself is the worst version of the problem, which is why the fix starts with a machine-readable About page.

How do engines resolve conflicting information about a person?

No lab publishes its exact reconciliation logic, so treat any confident percentage here as fiction. But the observable behavior of retrieval-backed systems points to three axes that decide which version of a fact wins:

1. Corroboration count. The version repeated by more independent sources tends to be treated as current truth. This is the same corroboration machinery behind the Network signal, covered in The Network Signal.
2. Source weight. A fact on an authoritative, frequently-crawled page outranks the same fact on a scraper site. Structured assertions help too: a schema.org Person declaration on your own domain, echoed by consistent profile pages, gives resolvers an explicit spine to reconcile against, and Google documents the pattern in its profile page structured data guidance.
3. Recency. Fresher assertions generally beat stale ones in retrieval, though facts absorbed into training data can resurface long after you fixed the live web. The two pathways age differently: retrieval forgets fast, training data forgets slowly.

The practical conclusion: you cannot argue with the weighting, but you can win on all three axes simultaneously. Make the correct version the most repeated, best-placed, most recent version in existence.

Scoring your own contradiction debt

Not every contradiction deserves equal urgency. Before you can triage a ledger of ten or twenty conflicts, it helps to score each one the way you would score technical debt in a codebase: by how visible it is, how authoritative the page carrying it is, and how recently it was asserted. Multiply a rough 1 to 3 rating on each axis and you get a simple debt score that tells you what to fix this week versus what can wait.

A simple 1-3 multiplier score. Anything above roughly 12 is a Grade A fix-this-week item; below 4 can wait for a quarterly sweep.

This scoring is deliberately rough. It is not meant to survive academic scrutiny, it is meant to stop you from spending an afternoon fixing a dormant social account nobody reads while a cited directory keeps handing engines your old title. Score the ledger once, sort by debt score, and work down the list.

The contradiction audit: seven steps

1. Define the canon. Write the single correct version of your facts: canonical name, title, company, niche descriptor, locations, credentials, dates. One document. This is your source of truth for everything that follows.
2. Sweep search. Search your name in quotes, plus each name variant, plus pairings with every past employer, title and city. Go past page one. Log every page that states a fact about you.
3. Interrogate the engines. Ask ChatGPT, Perplexity, Gemini and Claude who you are, what you do, and where you work, in multiple phrasings. Record every returned fact and its cited source. A structured version of this exercise is The 25-Prompt Audit.
4. Build the conflict ledger. One row per contradiction: the URL, the wrong fact, the correct fact, who controls the page, and a severity grade. Grade A: wrong on pages engines cite. Grade B: wrong on pages you control. Grade C: wrong on low-authority flotsam.
5. Fix what you control first. Your site, your profiles, your bylines' author boxes. This is a day of work and it removes the most embarrassing category: self-contradiction.
6. Request the rest. Email conference organizers, directory owners and editors with the exact replacement text, pre-written, so saying yes takes them ninety seconds. Most comply. Log who does not.
7. Re-test on a cycle. Re-run the engine interrogation quarterly and diff the answers against your canon. New contradictions appear whenever your career moves; the audit is a loop, not an event.

A worked example: one consultant, eleven contradictions

A composite from real audits, details changed. A supply-chain consultant, fifteen years in, ran the sweep and found eleven conflicts. Three were Grade A: a widely-cited industry directory listed her as "logistics manager" at a firm she left in 2021, a podcast page spelled her surname with a hyphen she does not use, and a university speaker archive credited her with a competitor's book. Five were Grade B, all on properties she controlled: two bio variants on her own site, an outdated LinkedIn headline, a Twitter bio naming the old firm, and a media kit PDF with a dead email address. Three were Grade C scraper copies of the old directory entry.

The Grade B fixes took one afternoon. The directory and the podcast page were corrected within two weeks of a polite email containing ready-to-paste replacement text. The university archive never replied, so she added a dated career timeline to her About page and let recency and corroboration outvote it. Two months later, the chatbot answers had converged on the current title, and the hyphenated surname had stopped appearing. Nothing in that story required special access or budget. It required a ledger and a fortnight of follow-through.

Why fixing your own site comes first, structurally

It is tempting to chase the highest-authority external contradiction first, since it feels like the biggest win. Resist that instinct. Your own domain plays a different role than every other source: resolvers treat it as the canonical reference point against which everything else gets checked, which is exactly why a machine-readable about page matters so much. If your About page, your footer bio, and your media kit disagree with each other, you have not just left a gap, you have removed the reference point a resolver would otherwise use to judge every external source.

Think of it as the difference between a typo in a witness statement and a typo in the court record. External contradictions are noisy testimony an engine has to weigh. Self-contradiction corrupts the ruler you were hoping it would measure everything else against. This is the mechanical reason the seven-step audit above puts your own properties first: not because they are easiest, though they are, but because until they agree with each other, no external correction has a stable target to reconcile toward.

Knowledge graph debt versus retrieval debt

Not all contradiction debt behaves the same way, and treating it as one undifferentiated pile leads to wasted effort. Split it by where the wrong fact actually lives, a distinction explored further in knowledge graphs for people.

• Retrieval debt lives in ordinary crawlable web pages: bios, directories, press. It is volatile. Fix the source page, wait for a recrawl, and within days to weeks the wrong version stops surfacing in retrieval-backed answers.
• Knowledge graph debt lives in structured, curated records such as a Wikidata entry or a Google Knowledge Panel. These are edited on a much slower cadence, sometimes requiring a manual claim or edit process, and once cached they can persist for months regardless of how fast the rest of the web moves on.
• Training data debt lives inside the model's weights themselves, absorbed at training time. It does not respond to any live-web fix at all. It only clears when a newer model version trains on more recent, corrected data.

The practical implication is patience calibrated to the layer. Fix retrieval debt and expect a fast payoff. Fix knowledge graph debt and expect a slow, sometimes manual, but eventually durable payoff. Training data debt is the one category you cannot accelerate, so the only defense is making sure the correct version is dominant and well corroborated well before the next training cutoff, so that whenever the next model does train, the debt does not get baked in again.

The repair ladder: when you cannot edit the source

Some contradictions will not die on request. Escalate in order:

• Update: the owner corrects the page. Best outcome, always ask first.
• Retire: the page or profile gets taken down or deindexed. Right answer for abandoned accounts you still control.
• Contextualize: where the old fact must stand (archived press, for instance), make sure your own properties state the change explicitly: "From 2019 to 2022, she led X; since 2023, she runs Y." Timelines convert contradictions into history.
• Outweigh: for scraper sites and immovable pages, stop fighting the page and win the reconciliation instead: publish and earn enough fresh, authoritative, consistent assertions that the stale version loses on corroboration, weight and recency all at once. This is where a running digital PR pipeline doubles as debt service.

A note on data aggregators specifically: many rebuild their records from upstream feeds, so a correction you win directly can be silently overwritten at the next sync. When an aggregator keeps regressing, find the upstream source it mirrors and fix that instead. Chasing mirrors is how audits turn into hobbies.

One special case deserves its own playbook: when the "contradiction" is actually another human with your name absorbing your facts, you have a collision problem, not a debt problem, and the disambiguation moves are different. That scenario is handled in Name Collisions.

Staying out of debt: the change protocol

Debt prevention is a protocol, not a virtue. When any fact about you changes, update in strict order within one week: your About page and site bios first, then your top three profiles, then your recurring bylines and speaker pages, then everything else in the ledger. Keep the canon document current, keep the conflict ledger alive, and treat every new placement's bio line as a deployment: it ships the canonical text or it does not ship. Identity consistency is unglamorous, which is exactly why it separates the recommended from the merely competent. The foundational version of this discipline is laid out in Structure Your Identity for Machines, and if you would rather have the audit run for you, that is on the services menu.