Separating a Reputation Event From Everything Else

Fix the date before you open the data

The analysis ahead compares a period before something against a period after it. Everything downstream depends on which day divides them, and that day is very rarely the day the content was posted.

Content appears, and then, separately, it starts being seen. A post can sit unread for eleven weeks before it surfaces on searches for the company name; an article can be picked up in August by a site that outranks the original. Splitting the data on a publication date, when the thing was invisible for a quarter, sets two periods against each other that have nothing to do with one another.

So fix the date from evidence rather than from the timestamp on the post:

• the first day the content shows up on searches carrying your own name, from your own search performance data;
• the first referral traffic from that source, if it sends any;
• the first inbound mention — an email, a call note, a support ticket, an objection written into a deal record;
• the first dated capture anybody took of the page or the result.

Where those disagree, the disagreement is the finding: write down a window rather than a day, and say why. An analysis resting on the wrong date is worse than no analysis, because it looks careful, and careful-looking things travel — into a board pack, an insurance notification, a letter somebody sends on the strength of it.

Write down what else changed before you look at the chart

Order matters here for a reason that has nothing to do with tidiness. Once the dip has been seen it cannot be unseen, and every explanation gathered afterwards gets weighed against a conclusion already held. The remedy is cheap: assemble the list of everything else that moved in the window before looking at the outcome.

It is not a memory exercise: the things that matter were done by different people who each thought their change was routine.

• Marketing — budget shifts, bidding and creative changes, channel mix, email frequency, and the change history the ad platforms keep for you.
• Engineering — the deploy log. Releases, template edits, redirects, tag and consent changes, anything touching how pages render or how activity is counted.
• Sales — headcount, territory, quota, discount authority, and who left.
• Finance and operations — price list versions, supply constraints, a large account ending, a checkout change.

Date every entry. The list written before anyone looks at the numbers is a control on the analyst rather than on the data, and it is the one step in this whole process that cannot be done halfway.

Seasonality is the cheapest check and often the whole answer

Compare the same weeks of the previous year, and the year before that — not the previous quarter, which is what people reach for because it is already on the screen and which answers a different question.

Several distinct things hide inside the word seasonality, and each has produced a confident false result somewhere:

• The composition of the days. Five Mondays against four, a public holiday falling midweek instead of on a Friday, a leap year. On a small base these move a monthly total by more than most reputation events do.
• Holidays that move. Easter, the week of Thanksgiving and where the New Year sits against a fiscal calendar all drag week numbers around, so week 14 is not week 14.
• Buying cycles. Quarter-end budget flush, renewal anniversaries, procurement freezes, academic calendars.

Two prior cycles is the minimum here and three is better. A business three years old cannot run this check honestly, and that limitation belongs in the write-up rather than being worked around. Where the same shape appears in the same weeks of two earlier years, the honest conclusion is that much of the decline was arriving regardless — a conclusion nobody commissions and everybody should want before spending money on the alternative.

The changes you made to yourself, and the question to ask first

Begin with the least welcome question, because if the answer is yes then everything after it is wrong: is this content a cause, or a record of something that was already going wrong?

There is a way to test that which does not depend on anyone's opinion of the business. Pull the operational series that would move first if service had actually slipped — refund and return rate, support ticket volume, response and delivery times, complaints arriving through your own channels, churn — and line them up against the content date. If those were deteriorating in the months before the reviews arrived, the reviews are downstream of the problem, and an analysis blaming them collapses the first time a competent reader sees the same series.

Then the deliberate changes, each of which moves conversion without touching demand: a price rise, a change to packaging or minimums, a shipping threshold, a shortened trial, a new required field in the enquiry form. Paid media leaks sideways — brand searches are partly manufactured by advertising, so cutting spend reduces the branded volume a reputation analysis then reads as people losing interest.

What the competitors and the category did

A competitor's decisions never appear in your analytics and always appear in your close rate, which makes them the confounder most often left off the list.

What is worth checking, and all of it is public: their pricing, whether they have started bidding on searches for your name, a funding announcement followed by a campaign, a new entrant, a distribution deal that put them in a channel you rely on, and their hiring, which telegraphs the rest by a quarter. Searching your own company name logged out once a week, and keeping the dated captures, answers several of these at once.

Then the category. Was the whole sector soft? Suppliers, trade bodies and the published results of any listed competitor are the cheap sources, and if a public company in your category reported the same shape in the same quarter, you have an external comparison you did not have to construct.

One caution about the tool everybody reaches for. A search interest index is a normalized series, not a count: it cannot be converted into a number of searches, and two pulls taken on different days can return different values for the same period. Useful for the shape of a curve, unusable as a figure in a document — the opposite of how it is usually presented.

The site change made in the same week

This is the most common false positive in the entire subject, and it is not close. Something is always being changed on a website, and shipping one in the same fortnight as the content appears is the fastest way to manufacture a convincing decline.

Two families of change, and they fail differently. The first alters what visitors can reach: a migration, a redesign, a navigation change, a redirect chain, a canonical or robots directive that quietly took pages out of eligibility. The second alters only the counting: a tag that stopped firing, a consent banner deployed to a new region, a bot-filtering change, a switch of attribution model, a metric whose definition moved underneath a dashboard. The second family is more common, harder to spot, and draws the same shape on a chart as a collapse in demand.

There is a straightforward test, and it belongs before anything else is concluded. Does the decline appear in a second series collected independently? Orders in the order system. Invoices in finance. Calls in the phone system. If the fall exists in analytics and not in the invoices, what you have is a measurement incident, and further analysis of the broken series will say nothing true.

When the instrument moved and the world did not

The platforms supplying the data have bad days as well, and at least one of them publishes them:

“On rare occasions, there might be an event in Search Console that could affect your report data.”

— Google Search Console Help, “Data anomalies in Search Console”, read 12 August 2026

Behind that sentence sits a running, dated log of those events, and they are neither rare nor small. One entry announced on 3 April 2026 concerns impression reporting across a stretch running from 13 May 2025 to 27 April 2026 — longer than most comparison windows get built. Other 2026 entries cover FAQ rich results ceasing to appear from 7 May 2026, with the drop in impressions that followed.

Reading that page takes two minutes and almost nobody does it. Two habits belong with it: the last few days of any window are incomplete rather than falling, because the data arrives with a lag; and applying a filter changes what the totals count, so a filtered period set against an unfiltered one produces a difference belonging to the filter. Google revises these pages without notice, so check the current version rather than a remembered one.

A control is something you choose before you know the answer

A comparison period picked after you have seen where the dip sits is not a control. It is the dip, selected. The test is simple and worth running on your own work: move the window four weeks in either direction. If the headline figure changes materially, it was a property of the window rather than of the business.

A real control is a series that was not exposed to the event and that moved with the affected series before it. Both halves matter and the second is the one people skip: if two lines did not track each other in the year before, nothing about the way they diverge afterwards means anything.

Candidates: a second location, a second product line, a segment the content never reached, a sibling brand under different naming, or a renewal book of customers who were never going to search for you. The comparison then runs on the difference between two changes rather than on one line falling, which removes whatever struck both.

It is worth seeing once what a designed comparison looks like. Michael Anderson and Jeremy Magruder, publishing in The Economic Journal in September 2012, studied restaurant reservation availability in a single city between July and October 2010, using the fact that a review platform rounds the star rating it displays: businesses just above and just below a rounding boundary are near-identical apart from the number a customer sees. The comparison was fixed by the design before any outcome was known, which is the entire point. The limits matter before borrowing authority from it — restaurants, one city, data now well over a decade old, and a star rating rather than an article. Work published with its method exposed, by authors with no outcome riding on the result, is a model for how a control gets built rather than a source of a number about your business.

The evidence that beats every chart

One email from a buyer saying they read the thing and went elsewhere outweighs any amount of time-series work, for a plain reason: it ties the content to a decision made by a named person, which is a different category of evidence from two lines moving in the same month.

Material like that exists only if somebody catches it while it is happening. Four ways to catch it, all cheap:

• A required field on lost deals asking whether the buyer raised anything they had read.
• A neutral question in the closing conversation with a lost prospect: whether anything they read or heard gave them pause. Do not name the article — naming it broadcasts the thing to someone who may never have seen it.
• Keeping the inbound messages that raise it, with headers and dates, rather than pasting the text into a summary.
• Logging consequences on other fronts in the week they occur — a candidate withdrawing, a procurement question, a lender following up.

Then the limit, in the same breath. A handful of documents proves the mechanism and says nothing about its size. Ten emails establish that people read it and acted; they do not establish what share of buyers were affected, because the ones who quietly never made contact are not in the file. Mechanism and magnitude are separate questions, and the analyses that survive scrutiny answer the first squarely and stay careful about the second.

What to write when the answer is that you cannot tell

Sometimes too much moved at once. A price rise, a migration, a soft quarter and a complaint page inside the same six weeks is not unusual. When that is what happened, the finding is that the effect cannot be isolated, and writing that down is the professional outcome rather than the failed one.

A report that survives a skeptical reader has four parts: the date and how it was established; what was observed on property you control, with sources; every competing explanation considered, each marked as excluded, still live, or untestable; and a conclusion pitched at the strength the evidence supports. In practice the strongest sentence available is that the timing fits, the alternatives capable of being tested were tested, and these particular ones could not be.

Two things do not work, and both are common. Running the analysis repeatedly with different windows and keeping the largest number is not analysis, it is selection, and anyone checking finds it within minutes by shifting the window back. And no analysis, however careful, reaches the content: nothing established about the cost persuades an operator to remove a review, changes what a search engine displays, or alters a word of what was published. What the work produces is a record — dated, sourced, and answerable when the question comes back around.

Frequently Asked Questions

Published 2026-08-12