Brand reputation metrics: six signals you can trust
Brand reputation metrics defined so they can be tracked without fooling yourself: SERP mix, reviews, replies, sentiment sample, AI accuracy, complaints.
Brand reputation metrics are numbers computed from dated observations of what the public can see and say about your company. Six hold up in practice: the mix of owned, third-party and negative results for your brand name, review volume, average and recency per platform, reply rate and time to reply, a sentiment reading from a fixed coded sample, an accuracy score for AI answers, and counts of complaint and regulator records. Each needs a written definition, or it drifts until it flatters you.
The danger is a number that moves for reasons unrelated to reputation: a different search location, a new sampling habit, a platform that changed what it counts. A trustworthy metric has a fixed query, sample and formula, and a stated blind spot.
This guide defines each signal so that two people measuring on the same day would get the same result. It includes a definition table, arithmetic on invented numbers and a list of vanity metrics to drop. Where a point rests on a help page, the page is linked and described as it read when checked.
- Track six signals: branded results mix, per-platform review stats, reply rate and time, a coded sentiment sample, an AI accuracy score, and complaint and regulator record counts.
- Write the definition before the first reading: query, sample, formula, rubric, cadence, owner. Change a definition only by starting a new series.
- Weight and normalize. A negative result at position three is not the same as one at position nine, and a complaint count means little without the orders behind it.
- Small samples lie. Report a margin with every sampled figure and call nothing a trend until the move is larger than the margin.
- Drop follower counts, raw mention totals and blended reputation scores.
What counts as a brand reputation metric and what does not
A brand reputation metric answers a question a manager could act on, and is computed the same way every time from something an outsider can see. A number that cannot be tied to what a customer, investor or candidate meets when they look you up is an activity count.
Three tests separate the two. Can someone else repeat it? If a colleague gets a different value, it is an opinion with a decimal point. Does it have a denominator? Complaints and replies need the orders or reviews behind them, or they only track growth. Does it have a stated blind spot? A results page is a snapshot for one place and device, and a review average ignores people who never reviewed.
Related readBrand Ambassador Scandal: The First 24 HoursReputation is also not one number. A brand can have excellent reviews and a damaging article on page one. That is why this guide keeps six separate series: a composite hides which surface moved.
Six signals in one definition table
The table sets out what each signal measures, its source, how often to read it and the failure mode that most often spoils it. Cadence is a practical default: read faster only when something is changing, and never slower than quarterly.
| Signal | What it measures and source | Cadence | Failure mode |
|---|---|---|---|
| Branded results mix | Position-weighted share of owned, neutral third-party and negative results for the exact brand name. Source: a logged manual search | Monthly | Different location, device or login each time |
| Review stats | Count, average and recent share, per platform. Source: each platform's page | Monthly | Merging platforms, or trusting an average on a tiny base |
| Reply rate and time | Share of reviews answered; median and slow-tail time to first reply. Source: your reply log | Monthly | Fast, generic replies that count as answered |
| Sentiment sample | Share positive, neutral, negative in a fixed sample coded to a rubric. Source: collected mentions | Quarterly | Choosing the sample after seeing it; one coder |
| AI accuracy score | Share of assistant answers that state a verified fact correctly. Source: fixed questions, repeated | Quarterly | One run treated as the answer |
| Complaint and regulator records | Complaints and register entries per 1,000 orders. Source: your log, public registers | Quarterly | Raw counts with no denominator |
Branded results mix: owned, third-party and negative
The branded results mix is the share of the first page for your exact brand name that you own, that third parties hold neutrally, and that is negative, weighted by position. A plain count hides where each result sits.
Fix the query: the exact brand name, one search engine, a private window, one country and language, one device type, and log the date. Label each of the top ten organic results as owned (your site and profiles you run), neutral third party (directories, press, partner pages), negative (critical coverage, complaint pages, hostile threads) or unrelated (another entity with the same name). Label by content, not hope: a thread whose top post is a complaint is negative, a neutral article with one critical paragraph stays neutral with a note. Write the rule once and apply it every time.
Weight by position. Give position one a weight of 10, position two 9, down to position ten with 1; the weights add to 55. A group's weighted share is its weights divided by 55.
Invented example. A company has owned results at positions 1, 2 and 7, negative results at 3 and 6, and five neutral results. A month later the owned results sit at 1, 2, 4 and 7 and the negative ones at 5 and 8. The count of negative results is two both times, so a count-based dashboard shows no change. Weighted, the negative share falls from 13 of 55 (23.6 percent) to 9 of 55 (16.4 percent) and the owned share rises from 23 of 55 (41.8 percent) to 30 of 55 (54.5 percent).
Related readEcommerce reputation management: what to fix firstComputed from the invented positions in the text. Real results vary by place, device and search history, so one reading is a sample, not a verdict.
A companion figure from your own data is branded search demand. Search Console's performance report defines clicks as the number of times a user clicked your site from Google Search results, impressions as how many times your site appeared, and click-through rate as clicks divided by impressions. Filter the Queries dimension to your brand name: 4,200 impressions and 1,260 clicks give 30 percent. The help page also warns that seeing a query in your list does not mean your site appears when you run that query yourself, because results vary by time, place, device and search history. See the Search Console performance report help.
Usual failures: searching signed in gives a personalized view, moving from top ten to top five mid-series changes the denominator, and relabelling a result once it becomes inconvenient hides the problem.
Review volume, average and recency, kept per platform
Review metrics are the count, the average and the share of recent reviews, reported separately for every platform. Never merge platforms: their populations, rules and counting methods differ, and a blended number describes no place a customer actually looks.
Google's help page on review scores says the score is the average of all ratings published on Google for that place or business, on a scale from 1 to 5 stars, and that a new review may take up to two weeks to show in the updated score (Business Profile help on review scores). The headline moves slowly because it averages all history, and it may lag, so log the date.
Add two derived figures per platform: reviews in the last 90 days, and their average. Invented example: a branch has 40 reviews averaging 4.5, which is 180 stars. Five one-star reviews arrive in a month: 185 stars over 45 reviews is 4.11. If seven of the earlier five-star reviews also fall in the last 90 days, those twelve reviews average (5 + 35) divided by 12, or 3.33. The headline looks acceptable; the recent figure predicts the next customer's experience.
Related readFounder reputation affects company: separate or alignComputed: 4.6 minus (4.6 times n plus 1) divided by (n plus 1), unrounded before subtraction.
Report the count beside every average, and set no target such as "reach 4.8" without a volume. Do not chase the number by asking only happy customers: the Business Profile help page on replying to reviews, linked in the next section, says reviews must reflect a genuine experience, and that offering free or discounted goods or services in exchange for reviews, changed reviews or removed negative reviews is strictly prohibited. Asking every customer at a consistent moment widens the base legitimately.
Reply rate and time to reply
Reply rate is the share of reviews that got a public reply from the company; time to reply is the interval from publication to your first reply. They show whether customers see that someone listens.
Define the clock: start at the platform's publication time, stop at your first reply, and state how weekends count. Report the rate for all reviews and for one-star reviews, because a high overall rate can hide silence where it matters. Use the median and a slow-tail figure, not the mean, which a few month-old reviews drag upward.
Invented quarter: 38 reviews, 29 replied, a rate of 76.3 percent. The 29 first-reply times in hours have a median of 15, a 90th percentile (nearest rank) of 75 and a mean of 28.2; 18 of 29 (62.1 percent) were answered within 24 hours. Most replies are quick and a few are very slow, which is the actionable finding.
Google's help page on replying says replies should be clear and helpful and short, should skip the generic thank-you sent to everyone, and should not be used to push deals. For negative reviews it says never to share a reviewer's private information or make personal attacks, and suggests acknowledging mistakes you made, a sincere apology, addressing the reviewer by name, referring to their feedback and responding promptly (Business Profile help on replying to reviews). A reply count alone therefore rewards five identical replies. Each month, check ten replies for a named issue, a next step and no private details.
Related readLayoffs and employer brand reputation: what stays publicSentiment sampling with a fixed sample and a written rubric
A sentiment reading is the share of positive, neutral and negative items in a small sample of public mentions, coded by people against a written rubric. It is slower than a dashboard and easier to trust, because you know what was counted.
Build it in five moves.
- Define the population: public reviews, forum posts, news and social posts naming the brand, in a stated quarter and in languages you can read.
- List the whole population before reading it, number it, and draw the sample by random numbers or every nth item. Fix the size in advance; sixty is enough to see large shifts.
- Write the rubric. Positive: the author states satisfaction or recommends. Negative: the author states a specific problem with the product, service, conduct or trust. Neutral: factual, a question, or no judgement. A mixed item is negative if it holds a specific unresolved problem.
- Have two people code independently.
- Compare, settle disagreements by discussion and record the result.
Invented example: two coders label the same 60 items and agree on 48, or 80 percent. Some of that is chance, so compute Cohen's kappa, which compares observed agreement with the agreement expected from each coder's label frequencies. With invented label frequencies that give an expected agreement of 0.348, kappa is 0.69. A low kappa means rewrite the rubric.
After reconciliation, 16 of 60 items are negative: 26.7 percent. At 95 percent confidence the normal-approximation margin for a sample that size is about plus or minus 11.2 points. Next quarter 17 of 60 are negative, 28.3 percent. A 1.7-point move inside an 11-point margin is not a trend. Report "no detectable change" and read the content of the negative items, which tells you what to fix.
AI answer accuracy score
An AI answer accuracy score is the share of assistant answers that state a verified fact about your company correctly, on a fixed question set. It measures what people increasingly meet instead of a list of links: one synthesized statement about your pricing, location, leadership or record.
Start with a fact sheet of verifiable statements, each tied to a source you control or a public record: what you sell, where you operate, leadership, price model, certifications. Write one neutral question per fact. Run each on each assistant you care about, more than once, because answers vary between runs, and store the full text with the date.
Related readLegal Dispute and Brand Reputation: A Company PlaybookCode each answer as correct, partly correct (right in outline, wrong or missing a material detail), wrong or no answer, and report the shares. Two people should code a subset. Invented example: 12 questions, 3 assistants and 2 runs give 72 answers: 51 correct, 14 partly correct and 7 wrong, or 70.8, 19.4 and 9.7 percent (rounded, totalling 99.9). The wrong answers cluster on two facts, and those facts are the work list. Also count questions where the runs disagreed; a fact right once and wrong once is not safely right.
The score cannot be compared across a changed question set, so add new questions as a separate series. Our AI search reputation page covers what we do about wrong answers; this metric only shows where to look.
Complaint and regulator record counts
These count formal complaints against you and entries in public registers that name the company, normalized by the business volume behind them. They are the least glamorous signal and often the most serious, because they leave a record outsiders can find.
Keep three buckets and never add them together: direct complaints your team logs, third-party complaint records on public sites and trade bodies, and regulator records such as upheld complaints, notices or licence conditions in an official register. Define what counts as one item: a complaint with three issues is one.
Then normalize per 1,000 orders, customers or transactions. Invented example: 18 complaints on 9,000 orders in a quarter is 2.0 per 1,000; the next quarter 27 on 10,500 is 2.57. The raw count is up 50 percent, the rate 29 percent. Both are worse, but the rate is the honest size of the change. Add the number still open and its age: a falling count with a growing pile of unresolved items is not improvement. For regulator records, read the register itself on a fixed date. Where a record is accurate and lawful, resolve the cause and add context where the register allows; do not look for a way to hide it.
Related readNew company online reputation with no reviewsVanity metrics, self-deception checks and what other work covers
A vanity metric rises with activity, not with improvement in what outsiders see. Drop these from the reputation report:
- Follower and subscriber counts. They measure audience, not trust.
- Raw mention totals. A spike can be praise, complaint or an unrelated name.
- Blended reputation scores. They hide which surface moved and cannot be audited.
- Impressions on your own posts. They describe your reach, not what a stranger finds.
Before any report goes out, ask: did the definition change, is the move larger than the margin, is there a denominator, would a skeptical reader accept the labels? Any problem goes in the report next to the number.
Four neighboring jobs are left to other work. A scorecard for an individual's name covers personal page-one exposure. A monitoring program covers alerts, thresholds and who owns a response, and feeds these readings. Software selection is a buying question; a tool's dashboard must still pass the same definition tests. The anatomy of the brand results page is its own topic.
A ninety-day timeline to a first trustworthy trend
You will not have a trend in month one. Fix definitions first, collect readings at the table's cadence, and read movement only when the move beats the margin.
- Days 1 to 10: write the definitionsName an owner per signal. Fix the query, location, device, sample size, rubric, fact sheet and denominators in one dated document.
- Days 11 to 20: take the baselineLog the results mix, platform review stats, last quarter's reply times and the complaint log. Prepare the sentiment population list.
- Days 21 to 40: code the first samplesTwo people code the 60-item sample and a subset of AI answers. Compute agreement and tighten the rubric if it is weak.
- Days 41 to 60: read the second monthly setRepeat the results mix and review stats. Write what is within the margin and what is not.
- Days 61 to 90: repeat the quarterly signalsRepeat sentiment, AI questions and complaint rates with the frozen definitions. Two points per series allow a first careful comparison.
Next step: write the definitions down
This week, write one page that fixes the six definitions, the owners and the first reading date. Reputation Partners can review the definitions and baseline with you, but the method works without us and the baseline is yours.
Start with the page for your exact brand name: it is the fastest signal to take and the first one outsiders meet. The free page one score gives you a structured way to read it before you set your own series.
Questions people ask
How many brand reputation metrics should a company track?
Six is a workable ceiling: results mix, review stats, reply behavior, a sentiment sample, AI accuracy and complaint records. More series usually means nobody reads them. Add another only when a specific decision depends on it, and retire any series that changed no decision in two quarters.
How often should reputation metrics be measured?
Match the cadence to how fast the signal moves. Review stats, replies and the branded results page suit a monthly read, because they change within weeks. Sentiment samples, AI answer checks and complaint rates suit a quarter, because shorter windows give samples too small to read.
Should we aim for a perfect review average?
No. A very high average on a small base can look implausible, and chasing it tempts people toward practices platforms prohibit. Aim for healthy volume, a steady flow of recent reviews, honest replies and a record of fixing the causes of complaints. Those are the parts of the number you can influence without gaming it.
What do we do when a metric looks bad?
First check the measurement: same definition, large enough sample, move beyond the margin? If it survives that, find the content behind the number, such as the negative result, the recurring complaint or the wrong AI fact, and route it to the right fix. Accurate criticism is answered by changing its cause, not by hiding it.
