Diagnosing a traffic drop, including the AI-era causes

Traffic falls and the first instinct is to start fixing things. That is what wastes weeks. A drop has a shape, and the shape identifies the cause faster than any audit. This page sets out the order to work in, the signature each cause leaves in the data, and what Google itself says recovery looks like.

Step one: establish what actually dropped

Before naming a cause, describe the loss precisely. Four questions cover it.

Clicks, impressions, or both? This distinction eliminates most of the cause list, and it is the one most reporting skips. It has its own section below.
Which pages? A drop concentrated on one template points somewhere technical. A drop spread evenly across the site points at ranking or demand.
Which queries? Losing your own brand queries is a different event from losing generic ones.
Which surface, country and device? Google's core update guidance recommends analyzing search types separately, because a drop can be confined to web results, images, video or news.

Dates matter as much as the numbers. Google's documentation on core updates recommends waiting at least a full week after an update finishes rolling out, then comparing that week with a week before the rollout started. Comparing an incomplete month against a complete one produces a drop that does not exist. Where these numbers live is covered in Search Console.

The cause list, with signatures and checks

Technical failure. Signature: impressions collapse rather than decline, often on one template or directory, sometimes within a day. Check: indexing status and HTTP status codes for the affected URLs, robots rules, canonical tags, and whether the pages still render. Start with technical SEO.
Manual action. Signature: a severe, abrupt loss on part or all of the site. Check: the manual actions report. Google's documentation states that "if a site has a manual action, some or all of that site will not be shown in Google search results" and that a reconsideration review is requested from the report once every affected page is fixed. Background in spam policies.
A core or spam update. Signature: the drop starts on a documented date, affects many pages at once, and shows up as position loss rather than as lost coverage. Check: the Search Status Dashboard, whose ranking history lists each named update with its start date and how long the rollout lasted. If your decline does not overlap a rollout window, it is not that update.
Seasonality. Signature: the same shape appeared last year. Check: compare year over year, never month over month. A category with an annual cycle produces a convincing fake crisis every time it turns.
Results page layout and AI features. Signature: impressions hold steady or rise while clicks fall. This is the pattern the Pew Research Center measured across 68,879 searches by 900 US adults in March 2025: on results pages carrying an AI summary, users clicked a traditional result on 8 percent of visits, against 15 percent where no summary appeared. See AI Overviews and AI Mode.
A lost featured snippet or rich result. Signature: a sharp click loss on a handful of high-volume queries while average position barely moves. Check: what the results page shows for those queries now, and whether you still meet the eligibility conditions.
Competitor movement. Signature: a gradual decline on a defined query set, with no site-wide pattern and no update alignment. Check: who occupies the positions you lost, and what they published or changed.
Changes you made. Signature: the decline starts on a release date rather than an algorithm date. Check: the deployment log. Migrations, redesigns, URL changes, robots edits, canonical changes and template rewrites are the most common cause of all, and the easiest to overlook because the change felt unrelated.

Impressions versus clicks, the AI-era distinction

These two losses have different causes and different remedies, and reports showing only sessions cannot tell them apart.

Falling impressions mean you are appearing less often. Something upstream changed: indexing, eligibility, ranking, or the volume of the queries themselves. That is a visibility problem and it belongs to the causes above.

Stable impressions with falling clicks mean you are still appearing and being clicked less. The results page changed shape around you, or the answer is now being resolved on the results page. In the Pew data, sessions ended on 26 percent of visits to a page carrying an AI summary against 16 percent without one, so a click rate that erodes while impressions hold is expected behavior in the categories where summaries appear, not evidence that your pages got worse. The response differs too: rewriting content does not recover a click that is no longer available, whereas being the cited source does.

Both rest on the crawl, index and serve foundation described in how Google ranks content, which is also where AI answers are grounded.

What recovery after a core update actually looks like

Google's core updates documentation is unusually direct, and it says less than most recovery advice implies. Core updates "don't target specific sites or individual web pages", and the page compares a ranking list to a list of favorite restaurants, where the entries that move down are not necessarily bad ones: there are simply other restaurants that now make the top 20. The premise that a drop proves a defect is not Google's premise.

On the size of the reaction, the documentation distinguishes a small position loss, where it says there is "no need to take drastic action" and recommends against changing content that is already performing well, from a large sustained loss, which calls for assessment of the site as a whole against the published self-assessment questions. It advises avoiding quick fixes, treats deleting content as a last resort, and warns that "there's no guarantee that changes you make to your website will result in noticeable impact in search results".

On timing: some changes can take effect within days, but confirming a site-wide improvement "could take several months", and if several months pass with no effect, that may mean waiting for the next core update. Any recovery plan promising a defined turnaround is promising something the source does not.

The honest close

Some drops are not failures. They are redistributions of demand: a query type that now resolves without a click, a category where the audience moved, a seasonal cycle doing what it always does. Those cannot be fixed, only worked around. Other drops are genuine defects, and those are usually cheap to repair once identified. Telling the two apart before touching anything is what the diagnostic order buys you.