# AI Answers About Job Candidates: Audit and Fix AI answers about job candidates: what assistants say when a recruiter searches your name, an eight-question audit, and how to fix the sources. --- AI answers about job candidates are built from whatever the assistant can retrieve or has learned about a name: public profiles, news, registers and any site the person owns. When a recruiter, hiring manager or investor asks about you, a small footprint or a shared name can make the answer thin, or mix you up with someone else. The practical response is to ask the assistants yourself, with the same questions each time, across several of them, and to write down what comes back. Then you make accurate sources easier to find and correct what is wrong at its source. This article covers the hiring case only. We make no claim about how many employers ask assistants about candidates, because we read no primary source that measures it.

The short version

An assistant answers a name question from retrieved pages, learned patterns or both. Small footprints and shared names are where answers go thin or wrong.
Audit with a fixed set of eight questions in at least three assistants, in fresh sessions, and log the date, the exact wording, the answer and the cited sources.
Make accurate sources easier to find: one owned profile page, the same name form on every profile, and structured data as the search provider documents it. None of it promises how an assistant will answer.
Employer-side rules differ by country. The UK and US pages we read say different things, and none was written about AI assistants.

## What AI answers about job candidates draw on An assistant answers a question about a named person from two kinds of material: pages it retrieves at the moment you ask, and patterns it learned in training. Which one dominates depends on the product and on whether web search is on for that question. OpenAI's help page for ChatGPT search says the product searches the web automatically when a question might benefit from it, and that responses using search carry inline citations and a Sources list. Google's page on AI features in Search says AI Overviews and AI Mode may use "query fan-out", meaning several related searches across subtopics and data sources, and that the responses and links shown vary between the two. Both pages describe the products as of the time read, and both products change. For a candidate, the reachable material falls into five groups. - **Your own pages.** A personal site, an about page or a portfolio. These are the only sources where you decide the wording. - **Profiles you control in part.** Professional networks, speaker pages, author pages and directories. You wrote them, but the platform decides what is indexed. - **Pages written by others.** News, trade press, event listings, podcasts, university pages, and team pages at current and past employers. - **Registers and records.** Company registers listing officers, professional registers with license status, and published regulator or court decisions. For a regulated profession these are often the heaviest sources. - **Aggregators.** People-search and data broker pages, which can be out of date or merge people who share a name. An assistant may quote a stale directory entry beside a current staff page with equal confidence, so the audit asks for sources as well as answers. ### Search on and search off give different answers With search, you see what the live web says about the name. Without it, you see what the model produced from training patterns, which is where invented details tend to appear. Record which mode each answer used, and where the product allows it, run a question once with search on and once with it off. OpenAI's page on how its models are developed says there is an element of randomness in how the model responds, so the same question may yield different answers, and that responses may sometimes include personal information about people whose information appears often online, such as public figures. One good answer proves little and one bad answer is not a verdict, which is why the audit repeats. ## Why small footprints and common names produce thin or wrong answers The two failure patterns need different fixes. **A thin answer** happens when there is little to retrieve. The assistant returns a short, generic summary built from one profile, or says it has nothing. That is not damaging in itself, but it leaves the first impression to a single page you may not have written well. **A conflated answer** happens when two people share a name. The assistant attaches a namesake's job, city or news item to you, because the name is the only key it has. Adding a role and a city to the question usually helps it choose, which is why the audit asks the bare question and then the same one with context. Three more patterns appear in candidate audits: stale facts, such as an old title repeated because the old page still ranks; unsupported specifics, such as a degree no source shows; and tone drift, where a neutral record is summarized with loaded words. Most wrong answers have a cause upstream: a page, a register entry or a missing page. Fix the cause and the answer follows when the assistant next retrieves. Only when the error sits in the model's own output is the provider's correction route the only route. ## The eight-question audit to run across assistants Run the same eight questions in at least three assistants, for example ChatGPT, Gemini, Perplexity or Microsoft Copilot. Use a fresh session for each, signed out where the product allows it, so your history does not shape the answer. Replace the brackets with your details and keep the wording identical.

No.	Question to paste	What it tests	Read the answer for
1	Who is [full name]?	The bare name, as a stranger would type it	Which person it picks, and whether it admits doubt
2	Who is [full name], [role] at [company] in [city]?	Disambiguation with context	Whether context removes the namesake
3	Summarize the professional background of [full name], [role], [city].	Career facts	Current employer, title, dates, education
4	What sources do you have on [full name], [role], [city]? List each with its link.	The source base	Which pages carry the answer, and which are stale
5	Is anything negative reported about [full name], [role], [city]? Cite sources.	The question a cautious reader would ask	Whether it invents, conflates or cites a real item
6	What licenses, registrations or memberships does [full name], [role], [city] hold?	Register facts	A match against the register itself
7	Is [full name], [role], [city] the same person as [namesake description]?	Namesake handling	Whether it separates the two or merges them
8	Write a three-sentence introduction of [full name], [role], [city] for a hiring manager.	The tone a reader gets	Fit, accuracy and loaded wording

Ask questions 5 and 7 only about yourself, and read the result as a test of the assistant. If you know of no namesake, skip question 7.

How to run the audit so results can be compared

Prepare the sheetOne row per assistant per question, with columns for date, assistant, search on or off, exact wording, answer, cited sources and score.
Open a fresh sessionUse a new chat for each assistant, signed out where possible, and paste each question exactly. A changed word is a changed test.
Save the evidenceCopy the full answer and every cited link into the sheet, and take a screenshot with the date visible.
Open the sourcesOpen every cited page. The page behind an answer is where the fix happens.
Repeat onceRun the set again a few days later. Differences between runs show how stable each answer is.

## How to score what you find Score each answer with one of five fixed labels, so a later audit can be compared with this one. - **Accurate.** The facts match your records and the cited sources support them. - **Thin.** Correct but short, generic, or based on a single page. - **Stale.** Correct for an earlier period, wrong now. - **Conflated.** At least one fact belongs to someone else. - **Unsupported.** A specific claim that no cited source shows. Two rules keep the scoring honest. A claim with no source is scored Unsupported even if it happens to be true, because a recruiter cannot rely on what cannot be traced. A claim you dislike that a source supports is scored Accurate, because the audit measures whether the answer matches the record, not whether it flatters you. Then find the cause in the cited pages. Thin usually means one page carries the answer. Stale usually means an old page outranks the new one. Conflated usually means a namesake page or an aggregator is cited. Unsupported with no source points to the model's own output, which no page change reaches directly. ## What to fix first: a decision table by candidate situation The four situations below cover most candidate audits. The first move differs, and so does the thing to avoid.

Situation	First move	What to avoid	Risk if ignored
Thin footprint	Publish one owned page with role, history and links to your real profiles, then complete the two or three profiles that already rank	Dozens of new profiles made only to fill space	One weak page defines the answer
Common name	Use one display form with a role qualifier on every profile, and state it in the page title and first line. Do not write about the namesake	A different form of your name on each profile	A namesake's facts attach to you
Old controversy	Check whether the item is accurate, current and complete. Add a factual, dated note on your own page only if it adds something true. Ask the publisher for an update where a basis exists	Trying to hide accurate public-interest reporting, which can draw a second article	The assistant repeats the item with no context
Regulated profession	Match your register entry to your owned page and profiles word for word, and link to the register	Claiming a status the register does not show	A mismatch reads as an inaccuracy, or worse

A short note on the old-controversy row. Accurate, lawful reporting of a public-interest event is not something this article helps hide. If an item is inaccurate, outdated or unlawful, the route runs through the publisher or platform, not the assistant. ## How to make accurate sources easier to find None of this makes an assistant answer a certain way, and none of it is gaming. It makes true facts easier to retrieve and harder to confuse. ### One owned profile page An owned page is the one source where you control every word. Put your display name in the title and first line, then role, employer, city, a short factual summary with dates, and links to your official profiles and any register entry. Keep it crawlable and indexable. Google's page on AI features says there are no additional requirements to appear in AI Overviews or AI Mode, no special markup and no AI-specific files, and that a page must be indexed and eligible to appear in Google Search with a snippet. It also says indexing and serving are not assured. So the ordinary work is the whole job: a clear page, crawlable, with facts that match your other profiles. ### Consistent profiles Check that every profile you control carries the same name form, role, employer and city. If you are "Jordan A. Reyes" on your page, do not be "J. Reyes" elsewhere. Close the profiles you no longer use, since a stale one can supply a stale fact. ### Structured data, as documented Google's ProfilePage documentation describes markup for pages about a single person or organization affiliated with a site, and names blog "About Me" pages and employee pages as valid examples. As read, the required properties are mainEntity, which is the Person, and name. Recommended ones include description, image, sameAs, alternateName and dateModified. The same page says Google does not promise that features using structured data will appear in results. Google's general structured data guidelines, also read, say markup must be a true representation of the page, must not cover content readers cannot see, and must not mislead, including by impersonating a person. So mark up what is visibly on the page, keep it current, and do not expect it to change an answer alone. ### What not to do - Do not create fake profiles, invent affiliations or buy mentions. Platforms can act against it, and an assistant can quote the result. - Do not delete a long-standing public presence in a hurry. A sudden gap can raise questions and the sources may persist elsewhere. If you want help building an owned page that holds up, our personal brand work covers it, and AI search reputation explains what assistants say about you more broadly. ## What employers are told about online information: UK and US pages Employer use of online material is shaped by data protection and consumer reporting rules. Each page below was read in this session and is limited to what it says. None was written about an assistant answering a name question, so how they apply to that use is for the employer's adviser to decide.

Country and source	What the page says	What it does not settle
UK: ICO, pre-employment vetting	Checking public social media profiles needs a justified, documented risk. Candidates must be told. Decisions based on online information alone are unlikely to be fair or lawful. Employers should let the candidate comment on accuracy	No mention of AI assistants. The page is a draft under review
UK: ICO, automated decisions in recruitment	Employers can use AI tools to assist decisions, but a person must stay meaningfully involved. Solely automated decisions with significant effect are restricted and need safeguards, including human intervention and a way to challenge	It covers tools used in decisions, not a recruiter's informal question. Also a draft under review
US: FTC, background checks for employers	The Fair Credit Reporting Act applies to reports from a company in the business of compiling background information. The employer gives stand-alone written notice and gets written permission, with further notices around adverse action	No mention of AI. The page text sets no rules for social media or search checks
US: EEOC, background checks for applicants	A 2014 publication with the FTC. Employers must treat applicants alike regardless of protected characteristics. A policy can be unlawful through disparate impact if not job related and consistent with business necessity. Reports can hold errors	Social media appears only in a list of topics an employer may ask about, with no search rules. City or state rules may also apply

On the UK side, both ICO pages say the guidance is under review because of the Data (Use and Access) Act and may change. The vetting page also says employers should not assume a candidate is untruthful when a check conflicts with what they said, and should let them explain. If an employer raises something found online, that guidance points to a chance to comment on its accuracy, so keep a dated screenshot and the correct source ready. On the US side, the FTC page says that when an employer relies on a background report, the person receives a copy of it and a summary of rights before the decision, and afterwards learns the reporting company's details and of the right to dispute it and get a free further report within 60 days. The EEOC publication says you can ask the reporting company to correct errors and tell the employer about the mistake. Whether a particular AI tool counts as such a reporting company is a legal question neither page answers, so do not assume either way. This is information, not legal advice. ## A worked example: one invented candidate, three assistants This scenario is invented. The person, employer and results are made up to show the method. **Assumptions.** Dana Whitfield is a fictional operations manager applying to a logistics company. Dana has a professional network profile, a former employer's staff page that still shows an old title, and no owned site. A different Dana Whitfield works in the same region in an unrelated field and appears in a local news item about a business dispute. Dana runs the eight questions in three assistants with search on, in fresh sessions: 24 answers. **Results as scored.** Ten Accurate, seven Thin, five Conflated, one Unsupported and one Stale. Ten, seven, five, one and one add up to 24. - The five Conflated answers come from question 1 (no context) and question 5 (negative reports) and cite the namesake's news item. With role and city added, most assistants separate the two. - The seven Thin answers rest on the network profile alone and repeat its wording. - The Stale answer gives the old title from the former employer's page. - The Unsupported answer names a degree in one assistant's reply with no cited source. **What Dana does.** Dana asks the former employer to update the staff page. Dana publishes a short owned page titled "Dana Whitfield, operations manager" with a line naming city and field and links to the network profile and the employer's current page, and uses that same name form on the network profile. Dana writes nothing about the namesake and creates no extra profiles. For the Unsupported answer, Dana records the date, assistant and wording and uses the provider route below, since no page change reaches it. ## A six-week plan and your next step The weeks below are our planning estimate for one person with a thin footprint, usually enough for this order of work and not a promise of any result.

Six-week schedule for a candidate audit and fix

Week 1: baselineRun the eight questions in three assistants, fill the sheet and open every cited source.
Week 2: causesName the page or missing page behind each non-accurate answer. Check register entries against your records.
Week 3: sourcesAsk owners of stale pages to update them. Align the name form and role across your profiles.
Week 4: owned pagePublish the owned page with visible facts and links, and add ProfilePage markup that matches the visible text.
Week 5: routesFor wrong or unsupported answers with no page cause, use the provider's route and keep the evidence.
Week 6: re-runRepeat the full set, compare with week 1, log the changes and decide what to monitor.

On provider routes, OpenAI's page says individuals in certain jurisdictions can object to how their personal information is processed, or make other data subject requests, through its Privacy Portal. It says requests are assessed case by case and it may not be able to fulfill one, for example if it cannot verify the information. We did not read other providers' routes, so check each provider's help pages first.

If this is happening to you now

If an assistant is giving a recruiter or investor a wrong or thin answer about you, keep the dated evidence and start with the audit above. If you want a second pair of eyes on the sources, Reputation Partners runs a private audit.

## Questions people ask ### Can I stop an AI assistant from answering questions about me? Not through one switch. OpenAI, for one, offers a Privacy Portal for people in certain jurisdictions, with requests assessed case by case and some declined. A successful request about one product does not reach other assistants or the web pages they retrieve. The steadier work is making accurate sources easy to find and correcting wrong pages at their source. ### What if the assistant says something that is simply false? Save the answer, date, assistant and exact question. Open any cited source, because the error may come from a page you can correct. If no source is cited, use the provider's correction or privacy route and attach the evidence. A provider may decline, so keep your own sources accurate in parallel. ### Does structured data help with this? It can help search engines understand what a page is about, and Google documents ProfilePage markup for pages about one person. That documentation makes no promise that features will appear, and the markup must match visible content. Treat it as tidy labeling of true facts, not a lever on what an assistant says. ### Do employers have to tell me they searched for me online? It depends on the country. The UK ICO vetting page, as read, says employers must tell candidates if they plan to use social media checks. The FTC page describes written notice and permission when a background report from a compiling company is used. Neither page was written about AI assistants, so ask the employer what it uses and why.