Using AI to sort and summarise client website feedback
A review round can leave you with forty comments from three people, some about the same button, some contradicting each other, and a few that quietly ask for a new page. An AI tool can summarize client feedback like this in seconds: group it by theme, flag what looks out of scope and draft replies. It can also misread it. This guide covers both.
Why sort feedback with AI at all
Reading every comment is not the slow part. The slow part is the sorting: noticing that three people asked for the same thing in different words, that two of them disagree, and that one innocent-looking note ("could we add a pricing page?") is new work. Then writing a sensible reply to each.
That sorting is pattern matching on text, which AI tools do well. Deciding what to do about it is judgement about your client, your budget and your relationship, which they don't. The workflow below gives the AI the first job and keeps the second one with you.
Step 1: get the feedback into one place, with context
An AI can only sort what it can see. A comment that says "make this bigger" means nothing without knowing which element "this" is. Before you start, gather every open comment with:
- the comment text and any replies
- who left it, so the AI can spot disagreements between people
- the page and the element it points to
- the screen size or view (desktop, tablet or phone)
- any suggested replacement text
In HTMLAnnotate, Copy for AI turns the open comments on a project into one brief with the element, CSS selector, view, page, suggested text and a screenshot link for each. You can also export comments as a CSV. If your AI tool supports MCP, it can read the comments directly with get_comments instead; letting your AI tool fix client feedback with MCP walks through that setup.
If feedback arrived by email or in a call, paste it in too, but mark it as such. It has no element attached, so the AI will have to guess where it applies.
Step 2: give the AI your scope
The AI can't flag out-of-scope requests unless it knows the scope. Paste the relevant part of your proposal or statement of work: the pages included, the number of revision rounds, and anything you listed as excluded. A short summary is fine. If you haven't written your round limit down, how many revision rounds to include covers how to phrase it.
Step 3: use a prompt that asks for structure
Ask for a fixed format so the output is easy to check. This prompt works in Claude, ChatGPT or most other chat tools. Paste your scope and the comments where shown.
You are helping a web agency sort client feedback on a website.
Our scope for this project:
[paste scope: pages included, revision rounds, exclusions]
Client comments (each with author, page, element, view and any suggested text):
[paste comments]
Please do the following, and don't change any comment's meaning:
1. Group the comments by theme (for example: copy, layout, images,
navigation, bugs). Under each theme, list the comments with their
author and element. Merge exact duplicates and say which you merged.
2. List any comments where reviewers disagree with each other.
3. Flag comments that may be outside the scope above. For each, quote
the comment, say which part of the scope it seems to go beyond, and
rate your confidence as high, medium or low.
4. List comments that are too vague to act on, and suggest one
question to ask the client for each.
5. Draft a short, friendly reply to each comment. For flagged or vague
comments, the reply should ask a question or say we'll confirm,
never promise the change.
Don't invent details that aren't in the comments. If you're unsure
what a comment refers to, say so.
Two parts of this prompt do most of the work. The confidence rating on scope flags tells you where to look hardest. The instruction to admit uncertainty makes it more likely the AI says "unclear" instead of quietly guessing.
Step 4: check these by hand
Don't skim the output and forward it. Go through these every time:
- Scope flags. Read each one against your proposal. The AI will miss things you agreed on a call and flag things you are happy to do. Every "this is extra" conversation is yours to have.
- Vague comments. "Make it pop" or "it feels off" can mean many things, and AI tends to pick one meaning and sound sure about it. Use its suggested questions as a draft, and see what to do with vague client feedback for the usual meanings.
- Merged duplicates. Check that two merged comments really ask for the same thing. "Bigger headline" and "shorter headline" are about the same element but not the same change.
- Disagreements. Decide who has the final say, or ask the client to. The AI shouldn't pick a winner.
- Tone and reading between the lines. A comment written in frustration, or a joke, can be read literally. You know the person; the AI doesn't.
- Every reply. Replies set expectations about time and cost. Read each one, and change anything that promises more than you mean to.
Step 5: send replies and make the changes
Once you have edited the replies, post them on the comments themselves so each answer sits next to the thing it is about. If you connected an MCP server, you can ask the AI to post the replies you approved and resolve the comments it has actually fixed, but say so explicitly: "Post only the replies I marked OK, and don't resolve anything yet."
For the changes themselves, a sorted, checked list makes a much better brief for an AI coding tool than the raw comments. Reviewing HTML written by ChatGPT or Claude covers the loop of sending fixes back and checking the new version.
A note on client data
Client feedback often contains more than design notes: names and emails, unreleased products, prices, internal politics. Before you paste it into an AI tool, check three things:
- Your contract. Some contracts and NDAs limit which third-party services can see client material. If yours does, follow it, or ask the client.
- The tool and plan. Use an AI tool your agency has agreed to use for client work, and look at its data and training settings.
- What you paste. Remove anything the sorting doesn't need, such as email addresses or confidential figures.
If you connect an AI tool to a feedback tool through MCP, the same rules apply: it can read whatever comments you give it access to. What an MCP server is includes what to check before you trust one.
What AI is good and bad at here
| Usually good at | Often gets wrong |
|---|---|
| Grouping comments by section or theme | Vague comments with no element attached |
| Spotting duplicates worded differently | Sarcasm, jokes and frustration |
| Flagging requests that clearly go beyond a written scope | Anything agreed outside the written scope |
| Drafting polite, consistent replies | Knowing what you can actually deliver, and when |
| Turning a sorted list into a change brief | Deciding between two reviewers who disagree |
Used this way, AI takes the sorting off your plate and leaves you the decisions, which are the part your client is paying you for.
Questions
Can AI summarize client feedback accurately?
For clear, specific comments, usually yes. It is good at spotting duplicates and grouping notes by section or theme. It is much less reliable with vague comments like "it feels off", sarcasm, or notes that depend on a call you had with the client. Treat its summary as a first sort, not a final list.
Can AI decide what is out of scope?
It can flag what looks out of scope if you give it the scope in writing. It can't know about promises made on a call or a goodwill change you are happy to make. A person should make every scope decision and have that conversation with the client.
Should I let AI send replies to the client?
Let it draft them, then read each one before it goes out. Replies set expectations about time, cost and what will change, so the wording needs to be yours. If you use an MCP server, tell the AI not to send anything until you approve it.
Is it safe to paste client feedback into ChatGPT or Claude?
Check first. Your contract or NDA may limit where client material can go, and comments can include names, emails, unreleased products or prices. Use an AI tool and plan your agreement allows, check its data settings, and remove anything sensitive you don't need.
What format should the feedback be in?
As much context per comment as you can get: the comment text, who left it, which page and element it refers to, the screen size, and any suggested wording. Comments pinned to the page give the AI far more to work with than a forwarded email.
Collect comments pinned to the exact element, then turn them into one AI brief with Copy for AI. Free, and reviewers never need an account.
Start a free review