What Siri, Alexa and Google Assistant say about you

What Siri, Alexa and Google Assistant say about you depends on the question and on the source the assistant reaches for: a web search result, a knowledge base, a business listing, a news article or a generative model. Each provider documents part of this and none documents all of it, so the honest answer is that a spoken reply about a person or a business comes from one of several upstream sources, and the fix is almost always at that source.

A spoken answer is one reply, usually a sentence or two, with no list of alternatives and often no visible source. That makes it hard to audit and easy to repeat.

This article separates what Apple, Amazon and Google say in their own help pages from what we infer, and sets out how to correct a wrong answer. Everything below is stated as of the pages read for this guide, and these products change often.

A spoken answer is one reply, and that changes how you audit it

A voice assistant gives one answer to one question, so you audit it by asking the question the way a stranger would and writing down exactly what came back.

Google's Pixel help page shows the direction in one documented case. For routines started by voice with Google Home connected to Gemini Apps, it says Gemini now responds with voice and text only and no longer displays visual information cards like maps or detailed weather graphs. That page is about routines, not questions about people, but the same logic applies to a spoken question about a person (our inference): the visual evidence you would use to judge a result is gone.

Related readAI Answers About Job Candidates: Audit and Fix

Two consequences follow.

• No ranking to read. You cannot see whether the reply came from the first result, a panel or a model.
• Variation. Generative assistants can word the same answer differently each time. Apple's help page on ChatGPT with Apple Intelligence says the generative models mean outputs may vary, and Google's Gemini help says Gemini may provide inaccurate or inappropriate responses about people.

Where each assistant takes its answer: documented and inferred

Each provider documents some sources, and the rest has to be inferred from behavior or from press reports, which the table marks as such. The table covers the four assistants in this article, on the surfaces they run on at the time of the pages read.

None of the pages read lists the sources used when you ask "who is" a named person, so treat any claim of an exact source, including ours, as describing likely routes.

Siri: a documented ChatGPT hand-off, a new beta, and an older search layer

Siri's behavior on questions about people is the least documented of the four, so this section keeps to what Apple's own pages say.

What Apple documents

Apple's guide to ChatGPT with Apple Intelligence says you are asked before any of your information is shared, and that ChatGPT uses generative models so outputs may vary. It advises you to check important information for accuracy. Apple's privacy page for the extension says it starts switched off and that nothing is sent until you turn it on.

Apple's page on getting Siri AI describes it as an entirely new version of Siri, in beta, that can go out to the web for up-to-date information to generate an answer. The page, dated October 22, 2026, says it is currently available in English, needs an Apple Intelligence-capable device on the version 27 software, and is not available on iOS, iPadOS or watchOS in the EU. It does not discuss errors or corrections and does not name the sources behind it.

Older search results and the business layer

A 2017 TechCrunch report, second-hand and nine years old, says Apple switched the default web search for Siri from Bing to Google, with Siri image results still from Bing at the time. No current Apple page read confirms it, so treat it as history. For businesses, Apple's help for Apple Business says a location's place card appears in many parts of Apple's software, from Maps to Siri. The page does not say how Siri uses it for a spoken answer. That part is inference, and a correct place card helps either way.

Related readAI answers about doctors and lawyers: what to fix

Alexa: a layered history, and an attribution habit that can misfire

Alexa answers have come from several layers, and Amazon's help pages that would settle the current mix could not be read for this guide, so this section separates what reporters found from what Amazon states now.

What Amazon states

Amazon's page on Alexa+, published on February 26, 2025 and updated on July 21, 2026, says Alexa's architecture rests on large language models available on Amazon Bedrock, infused with Alexa's knowledge base, and describes bringing information together into an accurate and real-time response. It reads as a product description, not a list of sources. It also says Alexa+ interactions are centralized in the Alexa Privacy dashboard, which is a sensible place to start when you want to see what was asked.

What reporters found

Earlier reporting describes the product at the time it was written.

• Crowd answers. In December 2018, TechCrunch reported that Amazon was opening Alexa Answers, an invite-only program, and that submitted answers would be noted as coming from "an Amazon customer." An undated TechHive piece later described it as open to Amazon customers, with downvotes leading Amazon to offer fewer of an answer. In November 2019 a VentureBeat analysis, known to us only through a Slashdot summary, said untrue and offensive answers were still accepted and that wrong answers stayed visible in its testing. All of this is second-hand, and we could not confirm from Amazon whether the program works the same way today.
• Reference and listing sources. A February 2019 Voicebot report found that some improved geography answers came from Wolfram|Alpha and some from Wikipedia, and that others gave no reference source. An undated Search Engine Land report said Echo local business data came from Yelp.

Attribution can be wrong

The strongest evidence that Alexa lifts text from web pages comes from the fact-checking charity Full Fact. On October 17, 2024 it reported that Alexa gave incorrect answers attributed to its fact checks, and said it suspected, without confirmation from Amazon, that Alexa had read the "claim" part of a check as if it were the answer. An Amazon spokesperson told Full Fact the answers were incorrect and that it was working to resolve the issue; Full Fact could no longer reproduce them later that day. On October 28, 2024 Full Fact reported more wrong answers attributed to other fact checkers and websites. Amazon said responses are pulled from a variety of sources, including licensed content providers and websites like Wikipedia, that these particular responses were pulled out of context, and that they were given a very small number of times. Full Fact found most had stopped, but one wrong answer attributed to an India-based entertainment news site was still being given on October 28.

Related readWrong company information in ChatGPT: how to fix

The lesson is narrow: a spoken "according to" line can name the right page and misstate it. Go to the named page and read it in full.

Google Assistant and Gemini: two products, one migration

The assistant on a Google device may be Google Assistant or Gemini depending on the device, the account and the date, so first find out which one answered. Google has said it is replacing Assistant with Gemini.

Phones

TechCrunch reported in March 2025 that Gemini would replace Assistant on Android phones. An undated PhoneArena piece quoted a Google support page saying Google had hoped to upgrade most mobile devices to Gemini by the end of 2025, and reported (its own summary) that the work moved into 2026. Google's Gemini help says that when Gemini is your primary mobile assistant it can do some of the same tasks as Assistant, that some familiar features may not be supported, and that devices without the Gemini app, such as Smart Displays, keep using Assistant.

Speakers and displays

Google's help page for Gemini for Home says that once you switch your home you cannot switch back to Assistant, and that older speakers get most features but not Gemini Live. It offers only a general helpfulness prompt, not a way to rate or correct an answer.

What Google documents about sources

Three documented routes matter for a person's name.

1. Gemini is cautious about people. Google's help on responses from Gemini Apps says it may provide inaccurate or inappropriate responses about people, so double-check them. It also warns that Gemini can misrepresent how it works, including when asked how it was trained, so we would not trust its account of where an answer came from.
2. News is read from publishers. Google's speakable documentation says Assistant, for news questions, can return up to three articles and read aloud sections marked speakable, attributing the source. The documentation limits this to U.S. users with Google Home devices set to English, calls it beta, and says eligibility is not guaranteed. If you are in the news, the spoken line may be a publisher's marked-up paragraph rather than your own page.
3. Knowledge panels are edited through verified feedback. Google's knowledge panel help says panels are generated from public web information, that verified people can suggest edits with supporting URLs, that Google reviews verified feedback within a few days though it can take more time, and that accepted changes are updated in the Knowledge Graph. The help does not say that spoken answers read the Knowledge Graph. We infer that they often do for simple factual questions, because the same company runs both, and we label it inference.

Why the answer is short, and why short is risky for reputation

A short answer hides the weakest link in the chain: a model or a ranking step that picked one claim and dropped the context. No provider documents these mechanics for spoken answers about people, so the three below are our inference from how voice queries work.

Related readGoogle AI Mode reputation: audit what it says

• Compression. An answer must fit in a few seconds of audio, so qualifiers go first. In an invented example with generic parties, "was sued in a dispute that was later dismissed" can become "was sued".
• Same-name merging. "Who is" plus a common name carries no disambiguating detail and can collapse several people into one answer.
• Stale or thin sources. An assistant that reads a single page will repeat it until the page changes. A business that moved, rebranded or changed hours is the usual case.

The providers put the burden of checking on the listener, so you need your own record.

How to audit what voice assistants say about you

Audit by asking a fixed set of questions on each device and account, and record the reply word for word. The set below takes about an hour and can be repeated monthly.

1. Write the question listInclude "who is" plus your full name, your name plus your profession, your name plus your city, and for a business "what does" plus the name "do", its address, its phone number and its hours.
2. Ask each device, onceUse the assistants you can reach: a phone with Siri, a phone with Gemini or Assistant, and an Alexa speaker or the app. Ask in a quiet room, the way a stranger would, with no follow-ups yet.
3. Record exactly what you hearSave the audio with your phone or a second device, then write it out with the date, device, assistant, account and language setting. Do not paraphrase.
4. Ask a follow-upAsk "where did you get that" only to note what it says. Google's help warns Gemini can misrepresent how it works, so treat the reply as a clue, not a citation.
5. Check your activity historyGemini Apps Activity lists your prompts at myactivity.google.com when Keep Activity is on. Per a TechRepublic how-to (undated), the Alexa Privacy web page lets you review voice history and grade a response. Do not rely on either for a record of the exact words.
6. Trace each claimSearch the claim text on a search engine, check the Knowledge Panel, your Maps or Business Profile listing, and any listing site. The page the claim matches is the source to fix.

Personalization can change an answer (our inference), so an answer on your own phone may differ from what a stranger hears. Ask on a device or account that has no history with you where you can.

How to correct a wrong spoken answer: fix the source, then report

Correct the page, listing or record the assistant read, then use the provider's route to report the reply. Reversing the order wastes the report, because the assistant will keep repeating its source.

For a person

Start with the claim and find where it lives.

• A publisher's page. If the line matches an article, the route is the publisher: a request to correct or update it, and, where the law supports it, a formal request. This varies by country.
• A knowledge panel or entity record. Verify your identity with Google and suggest an edit with supporting public URLs, as the help page describes.
• Your own profiles and a same-name problem. Make your site, bio and public profiles state the same facts in the same words, with a distinguishing detail such as profession and city, so a spoken answer has something to anchor to.

For a business

A business has more levers, because listings are structured data.

• Maps and listings. Google's Business Profile help says you can edit the profile directly on Search or Maps. Apple's Apple Business help says the place card appears from Maps to Siri. Neither page read explains how voice answers use the data, so that link is inference, and in our experience a wrong hour or phone number on these listings is a common cause of a wrong spoken business answer.
• Review sites and your own site. If an assistant reads a rating from a review site, use that platform's dispute route. Keep name, address, phone and hours identical everywhere.

Each provider's own feedback route

Each provider has a feedback channel, and each has limits that are documented, so use them with the right expectations. The table sets out what the pages read say, and where they say nothing.

Three notes follow from the pages. First, Apple's Report a Concern page says the tool works only while the response is on your screen, so capture the reply before you leave it. Second, Google's Gemini feedback page says feedback sends the full conversation, so do not submit a thread that contains other private details. Third, Google's Report legal issue option sits under More below a Gemini response. Take advice from a lawyer before you rely on any provider form as a legal notice. This is information, not legal advice.

Related readWhat Microsoft Copilot says about you: check and correct

A realistic timeline and what to expect

A source fix usually shows up in spoken answers in days to a few weeks, and no provider documents a clock for reports except Google's verified knowledge panel feedback. The ranges below are our experience, never a promise.

Expect uneven results: a speaker, a phone and an app may disagree for a while. If an answer is accurate and lawful but unflattering, the assistant may say it. The aim is a correct, current answer, never a hidden one.

If this is happening to you now

Our monthly monitoring repeats this check, and the page on what AI assistants say about you covers the written surfaces voice assistants borrow from. Reputation Partners handles the source fixes and the reports together.

Questions people ask

Can I ask Alexa or Siri what they know about me?

You can ask, and you will get one reply, but the reply is not a record of what the provider holds. Treat it as a sample. The sources behind it are the pages, listings and records that mention you, so the audit in this article is more informative than any single answer.

Does a voice assistant use the same sources as search?

Often, but the providers do not document it for questions about people. Apple, Amazon and Google each describe parts of their stack, and press reports describe search and reference sources. Where the pages read are silent, assume overlap with search but verify by tracing the claim, as the audit steps show.

Why does the answer differ between my phone and my speaker?

Different devices may run different assistants. Google is moving devices from Assistant to Gemini at different times, and Google's help says devices without the Gemini app, such as Smart Displays, keep using Assistant. Account settings, language and personalization also vary, so record the device with each answer.

Is a wrong spoken answer defamation?

It can be, but that depends on the law of the country, the claim and who is responsible, and a spoken reply from a model raises untested questions. A provider form is not a legal notice. If the claim is serious and false, speak to a lawyer in your jurisdiction and keep your recordings.

Where do I report a wrong answer from Gemini?

In the Gemini app, use the Good or Bad response button below the reply, or Send feedback under Settings and help. For a legal concern, Google's help shows a Report legal issue option under More below the response. Google notes the whole conversation is attached to the feedback, so review it first.