Reviews decide the shortlist: how to run a review program

Buyers build the shortlist before they ever contact you, and review sites are doing the work. A request cadence, response routing, and what badges actually get you.

Product marketing
August 18, 2026
·
6 min
read
Abstract brand form on a purple field

Your review profile gets read before your homepage does. A buyer who won't open a vendor site will happily open a category page, filter it to companies their own size, sort by most recent, and read six strangers describing what went wrong during onboarding. That's the first version of you they meet, and you didn't write a word of it.

Which would matter less if the shortlist were still forming by the time the sales call happens, and it isn't. 6sense's B2B Buyer Experience Report for 2025, built on nearly 4,000 responses, found that the point of first contact has moved to 61% of the buying journey, and that "ninety-five percent of the time, the winning vendor is already on the Day One shortlist". By the time you're in the conversation the list is written, and third-party sources wrote it.

So a review profile works like a reference check the buyer runs without telling you: you don't pick the referees or sit in on the call, and all you control is who you've made happy and how recently. That's the whole program, and it's deliberately unexciting: a steady rhythm between launches beats a scramble, so here's how we run one.

When should you ask for a review?

At the moment the customer is happiest, on a schedule you set once, and never two weeks before a badge deadline. Happiness is easier to time than it sounds. It clusters after a clean onboarding, after a support case that went well, and after a renewal, and each of those is already a field in your CRM.

The program we documented at Coro ran on four request channels, and the split between them mattered more than any one of them did.

ChannelWho it goes toWhen to run it
Automated campaignCustomers at 3 to 6 monthsAlways on
Automated campaignCustomers past 6 monthsAlways on
Platform-originated campaignThe review site's outreach listQuarterly
Manual note from the account managerNamed accounts onlyTime-sensitive

Look at what that last row costs, because a personal email from the account owner converts better than anything automated, which is why we keep it rare. Spend it on routine collection and you'll have nothing left for the quarter you fall short. Short of what, though? Badges carry a volume floor, and ours was 20 reviews a quarter minimum, which sounds steep until you divide it across a customer base and see a trickle you can automate. Miss the floor and the badge lapses, and that badge is on your homepage, in your deck, and inside your sales collateral.

Who asks, and who answers?

Different teams, on purpose. Customer Success drives collection, because they hold the relationship and know who's happy this month, while product marketing owns the responses, because a public reply is messaging read by people who are not the reviewer. So why not hand the whole thing to one team and save the coordination? Because collection rewards warmth and response rewards precision, and the people best at one are rarely who we want drafting the other.

Then there's the part nobody documents: money. Review incentives are real and they need a budget line, so we offered $20 to $25 per review, the first funded through the platform and later ones directly. Somebody had to top that card up every quarter, and the contract renewal is a second, larger conversation with finance that arrives annually whether anyone diarized it or not.

Does paying for reviews feel uncomfortable? A little, and we'd rather say so than pretend otherwise, because the platforms permit it, disclose it, and cap it, and what the incentive buys is fifteen minutes of writing rather than the opinion. If your customers only say good things when they're paid, you have a product problem, and no review program is going to touch it.

What do you do with a critical review?

Answer it. Publicly, quickly, and not in the voice of your legal department. Routing is what makes that survivable once volume climbs, so we build it before we need it:

  1. One channel. Notifications land in one place, email and chat, watched by a named person.
  2. Assign by topic. A coordinator hands each review to an owner who knows the product.
  3. Draft or delegate. The owner writes the reply, or finds the engineer who can.
  4. Check before posting. A peer reads it; anything sensitive goes to the marketing lead.
  5. Post, then route. Publish the reply, then route the substance to product, support, or the FAQ.

Step five is where a review program stops being marketing, because a critical review is unsolicited research from a customer who stayed long enough to have an opinion. And what about a review that's simply wrong? Answer that one too, with the correction and no adjectives, and let the next reader decide who sounds reliable. A wall of unbroken five stars reads as bought anyway, and one three-star review with a straight answer under it does more than the four glowing ones above it.

What do badges and profiles actually get you?

Less than the vendor's sales deck promises, and more than the cynics assume. Coro won G2 badges in every quarter I was there, from August 2022 to December 2024, including Grid Leader, Easiest Setup Mid-Market and Best Relationship, and Visual Layer, a computer vision company, was named in G2's 2025 icon refresh. Neither of them closed a deal for us, though both shortened the argument. The profile is the asset and the badge is only the receipt, so what we get alongside it is worth more:

  • Grid and index reports. Your category plotted with competitors, filterable to the two you meet in deals.
  • Comparison pages. Head-to-head data you can publish without writing a claim you'd have to defend, the logic behind the competitive comparison work case.
  • Quotable customer language, in the buyer's words, not yours. Keep it in one sheet, and get approval before attaching a full name.
  • Badges and matrices, refreshed quarterly.

That last one hides a trap, because badges expire quarterly and they live on the website, the slide template, the one-pager, the email footer, and a partner page somebody forgot. Every surface needs a named owner, or you'll ship an expired badge to the one buyer who checks the date.

What does a scramble look like from the buyer's side?

Obvious. That's the whole problem. Buyers sort by recency and filter by company size, so they see the shape of your collection rather than the total. Fourteen reviews inside one week, nine quiet months either side, and every one a variation on the same sentence: that's a campaign and it reads as one. Steady collection has no shape at all, which is the point: two or three a week, from different segments, about different parts of the product, some of them mixed.

One limit, since this is a method post. A review program won't fix your positioning, and it won't rescue a product people dislike. What it does is put an accurate version of you in front of a buyer who was never going to call you first, and that is now most of them.

Key takeaways

  • Ask on a schedule, timed to when customers are already happy.
  • Customer Success collects, product marketing answers, finance funds it quarterly.
  • Answer critical reviews fastest, and route what they say to product.
  • Give every badge surface a named owner, because badges expire.
  • Two or three a week beats forty in the fortnight before a deadline.

If your category page is thin and there's a launch in the calendar, bring it to a free 30-minute call. We'll work out who to ask and when, and you'll leave with a request cadence and a routing your team can run without you.

Mentoring, free forever.

I mentor marketers and technical writers. Breaking in, growing your career, automating with AI and more.

Book a free session
data-augmentation
Data Augmentation
Creating extra training examples by altering existing ones, such as flipping, cropping, rotating or recolouring images, so the model sees more variety.
Text Link
workflow
Workflow
The ordered steps and hand-offs by which work moves from start to finish, who does each one and what triggers the next.
Text Link
verification
Verification
Confirming that a result, label or component is correct by checking it against a reference or a second method.
Text Link
vectorized-execution
Vectorized Execution
A query-engine technique that processes whole columns of values in one CPU instruction instead of row by row, which is what makes engines like DuckDB fast.
Text Link
vector-space
Vector Space
The multi-dimensional space where embeddings live; distance within it stands for how similar two items are.
Text Link
uncertainty-sampling
Uncertainty Sampling
An active-learning strategy that sends the examples the model is least sure about to human reviewers first, because those teach it the most.
Text Link
transformer
Transformer
A neural network architecture built on attention, which weighs every part of the input against every other part; it underpins modern language and vision models.
Text Link
traceability
Traceability
The ability to follow any output back through the data, labels, code and model versions that produced it.
Text Link
tokenization
Tokenization
Splitting text into the units a language model works with, which may be words, parts of words or characters.
Text Link
token-limit
Token Limit
The maximum number of tokens a language model can take in and produce in one request, which caps how much context it can hold.
Text Link
threading
Threading
Running several parts of a program at the same time on separate threads so work overlaps instead of waiting in line.
Text Link
temporal-analysis
Temporal Analysis
Studying how things change over time across frames or records, to detect motion, events, trends or anomalies.
Text Link
sla
SLA
Service Level Agreement: the committed level of service a provider promises, such as uptime, response time or support hours, with consequences if missed.
Text Link
simulation
Simulation
Generating realistic scenarios in software to test or train a system where real data is scarce, risky or expensive to collect.
Text Link
similarity-search-vector-search
Similarity Search (Vector Search)
Finding the items whose embeddings sit closest to a query's embedding, which returns things that look or mean alike rather than share keywords.
Text Link
setup
Setup
The initial configuration of a project: connecting data sources, choosing models and settings, and defining the workflow before work begins.
Text Link
semi-supervised-learning
Semi-Supervised Learning
Training on a small labelled set plus a much larger unlabelled one, using the labelled examples to guide what the model learns from the rest.
Text Link
semantics
Semantics
The meaning behind data rather than its surface form, which lets a system treat "car" and "automobile", or two photos of one scene, as the same thing.
Text Link
segmentation
Segmentation
Dividing an image into regions by labelling every pixel, so the exact shape of each object or area is known.
Text Link
search-attributes
Search Attributes
The fields a search can filter or rank on, such as labels, tags, dates, camera id or model confidence.
Text Link
search-and-retrieval
Search & Retrieval
Finding the items in a collection that best match a query, whether by keyword, metadata or visual and semantic similarity.
Text Link
scene-understanding
Scene Understanding
Interpreting a whole image or video as a scene: what objects are present, where they are and how they relate.
Text Link
saliency-map
Saliency Map
A heat map over an image showing which pixels most influenced a model's prediction, used to explain and debug it.
Text Link
reward-function
Reward Function
In reinforcement learning, the rule that scores each action so the agent learns which behaviour to repeat.
Text Link
review-batch
Review Batch
A set of items grouped for a human reviewer to check in one sitting, often chosen because the model was unsure about them.
Text Link
relational-database
Relational Database
A database that stores data in tables with rows and columns and links them through keys, queried with SQL.
Text Link
regularization
Regularization
Techniques that stop a model memorising its training data, such as penalising large weights or dropping random neurons, so it generalises better.
Text Link
refactoring
Refactoring
Restructuring existing code to make it cleaner and easier to change without altering what it does.
Text Link
recipe
Recipe
A saved, repeatable set of processing steps or settings that can be rerun on new data to get the same kind of result.
Text Link
rag
RAG
Retrieval-Augmented Generation: a language model that first fetches relevant documents and then writes its answer from them, which cuts hallucination.
Text Link
quality-assurance-qa
Quality Assurance (QA)
The checks and processes that make sure a product, dataset or model meets its standard before it ships.
Text Link
predictive-maintenance
Predictive Maintenance
Using sensor data and models to forecast when equipment will fail, so it is serviced just before it does rather than on a fixed schedule or after a breakdown.
Text Link
policy
Policy
In reinforcement learning, the strategy that maps each state to the action the agent should take.
Text Link
pixel
Pixel
The smallest unit of a digital image, a single point with a colour or intensity value; images are grids of them.
Text Link
pgvector
pgvector
A PostgreSQL extension that stores embeddings and runs similarity search inside the database, so vector search needs no separate system.
Text Link
one-shot-learning
One-shot Learning
Learning to recognise a new class from a single example, usually by comparing against a learned notion of similarity.
Text Link
object-storage
Object Storage
Storage that keeps files as objects with metadata in flat buckets rather than folders, built for huge volumes and cheap durability.
Text Link
normalization
Normalization
Rescaling values to a common range or distribution so no single feature dominates and models train more stably.
Text Link
noise
Noise
Random or irrelevant variation in data that hides the real signal, from sensor error, compression, lighting or sloppy labels.
Text Link
mvcc
MVCC
Multi-Version Concurrency Control: a database technique that keeps multiple versions of a row so readers and writers do not block each other.
Text Link
monitoring
Monitoring
Continuously watching a deployed system's health and output quality so drift, errors and slowdowns are caught early.
Text Link
model-versioning
Model Versioning
Tracking each trained model with its data, code and settings, so any version can be reproduced, compared or rolled back.
Text Link
metrics
Metrics
The numbers used to judge a model or system: accuracy, precision, recall, F1, latency and others chosen to match the goal.
Text Link
metric-learning
Metric Learning
Training a model so that similar items end up close together in its embedding space and dissimilar ones far apart.
Text Link
meta-learning
Meta-learning
Training a model to learn how to learn, so it adapts to a new task from very few examples.
Text Link
mask
Mask
A pixel-level map marking exactly which pixels belong to an object or region, the output of segmentation.
Text Link
markov-decision-process-mdp
Markov Decision Process (MDP)
The mathematical frame for reinforcement learning: states, actions, rewards and transition probabilities where the next state depends only on the current one.
Text Link
loss-function
Loss Function
The formula that scores how wrong a model's predictions are during training; training works to make this number as small as possible.
Text Link
latent-feature
Latent Feature
A hidden characteristic a model learns internally that is not directly observed in the data but helps explain it.
Text Link
label-noise
Label Noise
Errors and inconsistencies in labels, from mistakes, ambiguity or disagreeing annotators, which degrade what a model can learn.
Text Link
keypoint-detection
Keypoint Detection
Locating specific points of interest on an object, such as joints on a body or corners on a part, rather than drawing a box around the whole thing.
Text Link
iou
IOU
Intersection over Union: the overlap between a predicted box or mask and the true one, divided by their combined area. 1.0 is a perfect match.
Text Link
inverted-index
Inverted Index
An index that maps each term to the documents containing it, which is how full-text search answers queries quickly.
Text Link
instance-object-instance
Instance (Object instance)
One individual occurrence of an object in an image, as distinct from the class it belongs to.
Text Link
information-extraction
Information Extraction
Pulling structured facts such as names, dates, amounts or relationships out of unstructured text or images.
Text Link
indexing
Indexing
Building a lookup structure over data so searches find matches without scanning everything.
Text Link
image-processing
Image Processing
Operations applied to pixels, such as resizing, filtering, thresholding or colour correction, usually to prepare images for analysis.
Text Link
hyperparameter-selection
Hyperparameter selection
Choosing the settings that control how a model learns, such as learning rate, batch size or network depth, which are set before training rather than learned.
Text Link
grid-search
Grid search
Trying every combination of chosen hyperparameter values in turn and keeping the set that scores best.
Text Link
gradient
Gradient
The direction and size of change in a model's error for each weight, used during training to decide how to adjust that weight.
Text Link
gpu
GPU
Graphics Processing Unit: a chip built for many parallel calculations at once, which makes it the workhorse for training and running neural networks.
Text Link
generative-ai
Generative AI
AI that creates new content such as text, images, code or audio from a prompt, learned from patterns in large training sets.
Text Link
generalization
Generalization
A model's ability to perform well on data it has never seen, rather than only on its training examples.
Text Link
feedback-loop
Feedback Loop
A cycle in which a system's outputs, or human reactions to them, are fed back as training signal so it improves over time.
Text Link
feature-map
Feature Map
The output of one layer in a convolutional network: a grid showing where in the image a particular pattern was detected.
Text Link
feature-engineering
Feature Engineering
Creating or transforming input variables so a model can learn from them more easily, using domain knowledge rather than raw data alone.
Text Link
fairness
Fairness
A model behaving without systematic disadvantage to groups defined by traits such as gender, age or region.
Text Link
exploration-vs-exploitation
Exploration vs. Exploitation
The trade-off in learning systems between trying new actions to discover better options and repeating actions already known to work.
Text Link
evaluation
Evaluation
Measuring how well a model performs against held-out data using agreed metrics, before and after it is deployed.
Text Link
efficiency
Efficiency
Getting the same or better result with less time, compute, money or manual effort.
Text Link
ec2
EC2
Amazon Elastic Compute Cloud: AWS's service for renting virtual servers by the hour, including GPU machines for training.
Text Link
domain-shift
Domain Shift
When the data a model meets in production differs from the data it was trained on, so accuracy drops even though nothing in the model changed.
Text Link
dml
DML
Data Manipulation Language: the SQL commands that read and change rows, such as SELECT, INSERT, UPDATE and DELETE.
Text Link
dino
DINO
A self-supervised vision method from Meta that learns strong image features without labels by teaching a network to match different views of the same image.
Text Link
ddl
DDL
Data Definition Language: the SQL commands that create and change the structure of a database, such as CREATE TABLE and ALTER.
Text Link
database-reliability
Database Reliability
A database's ability to stay available, correct and recoverable under load, failure and change.
Text Link
data-quality
Data Quality
How fit data is for its purpose: accurate, complete, consistent, current and free of duplicates, noise and wrong labels.
Text Link
data-lineage
Data Lineage
The record of where data came from and every transformation it went through, so any result can be traced back to its source.
Text Link
data-leakage
Data Leakage
When information from the test set, or from the future, slips into training, so the model scores well in the lab and fails in production.
Text Link
data-lake
Data Lake
A store that keeps raw data in its original format at any scale, structured or not, for later processing and analysis.
Text Link
data-integrity
Data Integrity
Data that is complete, accurate and unaltered over its whole life, protected from corruption, loss or unauthorised change.
Text Link
data-ingestion
Data Ingestion
Bringing data from its sources into a system where it can be stored, processed and analysed, whether in batches or as a live stream.
Text Link
data-governance
Data Governance
The policies, roles and controls that decide who can access data, how it is kept accurate and how its use is tracked.
Text Link
data-consistency
Data Consistency
Data that stays correct and in agreement across copies, transactions and systems, so every reader sees the same truth.
Text Link
data-balance
Data Balance
How evenly examples are spread across classes or conditions in a dataset; imbalance teaches a model to favour whatever it sees most.
Text Link
cross-validation
Cross-validation
Evaluating a model by splitting the data into several folds, training on some and testing on the rest in turn, to get a more honest accuracy estimate.
Text Link
cosine-similarity
Cosine Similarity
A score between -1 and 1 for how closely two vectors point in the same direction, used to compare embeddings regardless of their length.
Text Link
convolutional-neural-network-cnn
Convolutional Neural Network (CNN)
A neural network built for images: it slides small filters over the picture to detect edges, textures and shapes, then combines them into higher-level features.
Text Link
connected-components
Connected Components
Groups of items linked by similarity so that every member connects to the group through at least one close neighbour, used to find clusters and duplicates.
Text Link
compliance
Compliance
Meeting the laws, standards and internal policies that govern how data is collected, stored, used and audited.
Text Link
collaboration
Collaboration
People and teams working on shared data, models and decisions with a common view of the work, rather than in separate tools and silos.
Text Link
cloud-native
Cloud Native
Software designed from the start to run in the cloud, typically built as containers and services that scale and recover automatically.
Text Link
cloud-computing
Cloud Computing
Renting computing resources such as servers, storage and databases over the internet, scaling up or down on demand instead of owning hardware.
Text Link
cloud-ai
Cloud AI
Running AI training or inference on a provider's infrastructure rather than your own hardware, paying for compute as you use it.
Text Link
clip
CLIP
A model from OpenAI that learns images and text in one shared space, so a sentence can be matched to pictures it describes without task-specific training.
Text Link
caching
Caching
Keeping a copy of frequently used data in fast storage so later requests are served without recomputing or refetching it.
Text Link
buy-vs-build
Buy vs Build
The decision between purchasing an existing product and developing your own, weighing speed, cost, control and long-term maintenance.
Text Link
business-value
Business Value
The measurable benefit a project delivers to the organisation: revenue, cost saved, risk reduced or time recovered.
Text Link
black-box-model
Black-box Model
A model whose internal reasoning cannot be inspected or explained, so you can see what it predicts but not why.
Text Link
big-data
Big Data
Data too large, fast or varied for traditional tools to store and analyse, which calls for distributed storage and processing.
Text Link
Keep reading

Tell me what you're building and what's in the way. I read everything.