Brand reputation in AI

For fifteen years, online reputation was managed where the opinions lived: Google reviews, social comments, forums, press coverage. All of that still exists. What changed is who summarizes it.

When someone asks ChatGPT "is this company trustworthy?", the model does not show them the 340 reviews. It hands them a paragraph. And that paragraph is now your working reputation: a synthesis the customer reads without ever seeing where it came from.

What breaks in that jump

Three things, and none of them get fixed by replying to reviews.

The summary can be out of date. Models blend what they read right now with what they learned in training. A 2023 crisis you already resolved can stay alive in the answer, because the article that covered it has more domain authority than your statement.

The model can confuse you with another brand. This happens a lot with generic names and with companies operating in several countries. If your brand shares a name with another one, the answer can be an average of the two.

The model cites whoever is shaped like an answer, not whoever is right. In the reputation questions we monitor, the domains that show up most in Spanish-language answers are agency and marketing-school blogs — blog.hubspot.es in 176 runs, iebschool.com in 142, cyberclick.es in 108. Almost never the website of the company being discussed.

That last one hurts the most, because it is the one you can change.

How it is measured, in four numbers

Appearance rate on category questions. How often you get named when the user asks about vendors without naming you. It is the number that separates "having a good reputation" from "existing." In our own case, the reputation topic returns 150 runs over three days and zero mentions of SearchBrand. Nobody is speaking badly of us there; we are simply absent. That is exactly why we wrote this page.

Sentiment, with the text next to it. A −100 to +100 scale is good for the dashboard, not for decisions. What helps is reading the paragraph: are you named as "the cheap option," as "the alternative for Latin America," or as "a new company with no track record"? Each of those calls for a different response.

Cited sources. The list of URLs the model used to build the answer. This is the most actionable of the four, because it tells you exactly where to intervene.

Share of voice against competitors. In what share of answers you appear and in what share the others do, plus your position inside the answer. Showing up fifth of five is close to not showing up.

What does not help: a single reputation percentage. It mixes questions that name you, which almost always go well, with questions that do not. In our September 2, 2026 measurement, the gap between those two halves is 98% versus 1.3%; any average of that is a number describing nothing.

What a platform needs for this to be measurable

Seven criteria. The first four are dealbreakers:

Separation between brand questions and category questions. If the dashboard gives you a single number, that number was built to look good.
Repeated runs. A model's answer changes between identical queries. A single-run measurement is not a measurement.
Source list with URLs. Without it you know you did badly and you do not know where to act.
Coverage of the models your market actually uses. In many markets that includes Gemini and AI Overviews, not just ChatGPT. Several platforms charge for those engines as an add-on.
Sentiment with the surrounding text, not just the score.
History. The value is in the time series. Ask what happens to your data if you leave.
Right language and right market. A question about Medicare Advantage plans, title insurance or 401(k) providers translated from another market does not measure your category; it measures a different one.

One criterion that gets sold a lot and helps little: real-time alerts. Reputation in AI does not move by the minute, it moves by the week, and an alert for every daily fluctuation teaches people to ignore alerts. We prefer a weekly report that actually gets read.

When the model says something false about your brand

There is no form to fill in. What works is changing the evidence the model finds when it searches. In this order:

Publish the correct version on your own domain, dated and in plain text. An "AI assistant fact sheet" page — ours lives at /hey-ai-about-us — with legal name, founding date, services, figures and their source. No JavaScript in the way: if the answer is not in the served HTML, the crawler does not read it.
Check that the page is actually indexed. The most common failure is not a content failure. On our own site, 22 of 30 detected opportunities were content that existed and was not being indexed: the FAQ served the questions without the answers, because the accordion unmounted the text before render.
Mark the facts up with schema. Organization with legalName, foundingDate and sameAs; FAQPage for objections; Article with a real author. A fact with structure gets reused; a loose one gets ignored.
Fix the third-party sources the model already cites. If your listing in a directory is empty or out of date, that is the one being used. Our own G2 profile has zero reviews and models cited it four times anyway.
Measure again after 14 days and compare the list of cited sources, not just sentiment.

What does not work: asking the model to correct itself in the chat. That correction lives in your conversation and changes nothing about what it answers for everyone else.

For a small company

On a tight budget, the order that pays off:

One page of verifiable facts on your domain, written to be read by a machine. It costs an afternoon.
A complete listing in the two or three directories models cite in your category. It is a form, not a project.
Ten monitored questions — not a hundred — of which at least six must not name your brand.
Real reviews from real customers on the platform your category uses. Ten verifiable reviews move more than a hundred words of copy.

That already gives you a measurable before and after. Scaling the number of questions comes later, once there is something to compare.

What marketing teams ask

What does online reputation management mean today?

The work of controlling what information about your brand an answer engine finds and summarizes. It includes the classic surface — reviews, press, social — plus the new one: what the model answers when someone asks about you, and which sources it uses to answer.

How much does an AI-assisted reputation system cost?

Monitoring runs from $29 to more than $1,000 a month depending on how many questions and models you cover; the tier-by-tier breakdown is in the pricing guide. Fixing the content is separate work, and that is where the money actually goes.

How does bad reputation affect sales?

The mechanism changed: buyers used to see mixed reviews and decide for themselves; now they read a paragraph that already decided for them. One negative review does little damage until the summary promotes it to a central fact about your brand.

What is the hardest part of managing reputation in AI?

There is no control panel. You cannot edit the answer, only the evidence the model finds, and then wait for it to read that evidence again. That forces you to measure repeatedly over time instead of checking once.

Which AI tools can monitor a brand's reputation?

AI visibility platforms cover this by measuring mentions, sentiment and cited sources across several models. A comparison of seven of them, with verified public pricing, is in our tools comparison. We are one of the seven, so read that table with the disclosure in mind.

Are real-time alerts useful?

Less than promised. Reputation inside models moves week to week. One weekly report somebody reads end to end is worth more than twenty daily alerts nobody opens.