Notes on running a support surface that does not embarrass you

What the word actually means, and which questions are worth taking off the queue.

• Most of what people ask is asked once14 Aug 2026The tail of rare questions looks like a to do list and is mostly a distraction. The argument against writing for it, and the one kind that earns a page.
• Five conversations you should never let it handle16 Jun 2026A complaint, a decided cancellation, a person in distress, a legal or medical question, and a first sales conversation. The business case for each.
• The cheapest customer research on your own website4 Jun 2026Site search queries are visitor wording captured at the moment of failure. How to read the zero result rows, and why phrasing matters more than topic.
• Why a confident wrong answer costs more than a refusal11 Mar 2026A refusal costs one extra step. A wrong answer costs a wasted journey, a lost sale, or a complaint, and it discredits every correct answer around it.
• How to find the questions your customers actually ask24 Feb 2026A method for pulling real top questions out of your own inbox: read a week, group by intent rather than tag, count, then look for what already exists.
• Deflection is four different things and only one is good10 Feb 2026The word covers a ticket avoided, a question answered, a conversation that ended, and a customer who gave up. Telling them apart.

When an assistant should stop answering, and what has to happen next.

• Eleven at night, in five different trades30 Jul 2026One message, five businesses, five different right answers. What a letting agent, a vet, a venue, an electrician and a small shop each owe by morning.
• When a visitor is in distress on a support widget6 May 2026Why a support assistant should not try to help somebody in crisis, and why the behaviour must be a hard rule rather than a hope.
• Collecting an email address is not the same as reaching a person23 Apr 2026What becomes of a message that lands in an inbox nobody owns, the reply time a form implicitly promises, and how to word an expectation you can keep.
• The conversation almost nobody designs for21 Apr 2026Bereavement notifications arrive through ordinary support channels. Why the standard flow fails them, and what the assistant should say instead.
• A complaint answered automatically is a complaint made worse2 Apr 2026Acknowledge, route, set an expectation, and never accept or deny liability. Where a regulated escalation route has to be named by name.
• What to promise when nobody is there17 Mar 2026The difference between an emergency route and a message form, and what to write instead of we will get back to you shortly.
• Escalation triggers that belong in a rule, not a prompt4 Mar 2026A request for a person, a symptom, an allergy, a fraud report, a safeguarding disclosure, a legal threat, a bereavement. Why each needs a rule.
• Designing the handover: when an assistant should stop19 Feb 2026Offer a person early, never behind a maze, collect only what makes a reply possible, and promise a reply time you can actually meet.

Writing material an assistant can answer from, and keeping it true.

• The page that is correct and useless7 Jul 2026A policy written to survive a dispute cannot answer a question, and the friendly version written next eventually contradicts it. How to publish both.
• The paragraph is the unit, not the page30 Jun 2026Readers now meet reference writing one paragraph at a time. What that does to pronouns, to relative dates, and to sentences that lean on the one before.
• Which of your documents will be wrong first9 Jun 2026Prices, stock, hours, staff names and regulatory thresholds rot fastest. How to notice, and why pointers beat uploaded snapshots.
• How to write the message your assistant sends when it does not know21 May 2026The most read sentence in the whole system. What a good refusal contains, what a bad one does, and why apologising and stopping is the worst option.
• Your opening hours are published in five places and they disagree5 May 2026The most common content failure in small business support, why nobody notices, and why an automated assistant turns it from harmless into expensive.
• A table is a shape, and the shape is what gets lost28 Apr 2026Price grids, size charts and delivery matrices each break in their own way when flattened. How to write the sentence version without deleting the table.
• Audit your help content before you let anything index it14 Apr 2026Find the duplicate answers, the contradictions, the out of date prices and the questions with no page at all, before indexing turns them into answers.
• When the right answer depends on which version they are on9 Apr 2026A version selector splits one body of material into several. Treat an archived release as a second corpus, and make every dated answer state its date.
• Your most exact document is a PDF, and that is the problem7 Apr 2026What happens to a fee schedule, a manual or a scanned form when a retrieval system reads it, and how to choose between rewriting and indexing.
• The shape of a help page that can actually be answered from30 Mar 2026One question per heading, the answer in the first two sentences, nothing important trapped in an image, and conditions written out rather than implied.

The numbers that mean something, and the popular one that does not.

• Two months that are not comparable6 Aug 2026A redesign, a price change, a campaign and the season all moved between the two months you are comparing. Reading the difference without fooling yourself.
• An answer that shows its source is a different product5 Aug 2026Grounded answers are checkable, correctable and fail visibly. Ungrounded ones fail silently. What makes a citation useful rather than decorative.
• Was that answer actually right?14 Jul 2026A practical sampling approach to answer quality, why citations make checking cheap, and why the checker has to know the subject.
• How to read conversation logs without wasting an afternoon25 Jun 2026A method: read the refusals, the rephrased questions and the abrupt endings. The conversations that went fine teach you nothing worth the time.
• The containment rate trap9 Jun 2026An assistant that never offers a person contains every conversation. What to pair containment with so it stops being a lie.
• Every product in this category meters something different4 Jun 2026Conversations, messages, resolutions, seats, indexed pages, stored words. Why headline prices do not compare, and how to price your own volume.
• The metrics worth putting on a support assistant26 May 2026Refusal rate by topic, unmatched questions, handovers by topic and repeat asks. Why conversation counts and two-button ratings tell you little.
• How to read a sample so it produces something other than a feeling19 May 2026Draw the sample before you look at it, record the same fields on every conversation, and repeat on a fixed day. Why this beats a dashboard early on.

Ownership, policy, launch weeks, and running one on a site nobody visits yet.

• The one page policy to write before launch27 Aug 2026What the assistant may state as fact, what it must refuse, who may change the material, and how a wrong answer gets reported and fixed.
• When the owner is the queue27 Aug 2026Support automation advice assumes a team, a rota and a manager. What actually changes when one person answers everything and nobody can cover.
• The case for an assistant that only does one thing25 Aug 2026Scoped to one product line, one audience or one section of a site, an assistant is more accurate, easier to keep true and easier to explain.
• Running an assistant on a site that gets very little traffic4 Aug 2026What ten conversations can and cannot tell you, why a deflection metric is meaningless at this scale, and why it is still worth having.
• The pilot that never ends21 Jul 2026These projects rarely fail outright. They stall on a quiet page with no owner and no date. The decision that is missing is usually not about the tool.
• Nobody owns the assistant until it says something wrong14 Jul 2026Support wants accuracy, marketing wants conversion, legal wants caution, and it answers in one voice. A workable ownership model and who decides what.
• The refusal sentence nobody was asked to approve14 Jul 2026It states what the business will not answer, where to go instead and how fast. It gets typed into a settings box in a minute, by whoever installed the thing.
• What to do in the first week after switching an assistant on25 Jun 2026Read every conversation, start with the refusals because they are your content gaps, and change the material rather than the wording of the prompt.
• When to remove an assistant instead of improving it18 Jun 2026The signals that say it should go, how to take it down without leaving a broken page, and what the decision reveals about the material underneath.
• The first time it says something wrong in public29 May 2026What to do in the first hour, the first day and the first week, why the fix is nearly always in the material, and who owes the reporter a reply.
• When one assistant should not cover both your websites12 May 2026Brands, regions, departments and a shop on its own domain. Where the boundary between them belongs, and what a leak across it actually costs.
• The visitors who never see your assistant at all8 May 2026A consent layer can hold a third party script back before it loads. Who that hides your widget from, what your analytics will not tell you, and the options.

Deliberate. The navigation at the top of the site has one job, which is to get somebody who is evaluating the product to the thing they came to evaluate. Adding a reading list to it makes that job harder for everyone and easier for nobody.

These pages are reachable from the footer of every page, from each other, and from the sitemap, which is how anybody who wants them actually arrives.

44 pieces so far.