Claude Opus 5 won't change your pages — but it may move your numbers

Over the course of March 2026, the share of ChatGPT answers citing no source at all went from 28% to 48% in the United States. On April 19, citation volumes fell by more than 80% in the US, the UK and Germany. By May, they were climbing back toward where they had been.

Then, on July 24, Anthropic shipped Claude Opus 5 — and the question landing in your inbox this week is entirely predictable: do we need to change anything?

For your pages, no. For the way you read your numbers, yes — and that part matters more than it sounds.

The seven weeks when AI citations fell off a cliff

Start with the measurement, because it reframes everything else.

From February 8 to April 27, 2026, seoClarity tracked the average number of sources ChatGPT cited per prompt across five markets — the US, the UK, Canada, Germany and Italy — drawing on several million interactions. The analysis is by Mitul Gandhi, seoClarity's co-founder, last updated June 19, 2026. The chart reads like a heart monitor.

Look at the Italian line: at the moment three markets hit bottom, that one turns upward. That's what separates a tremor from a trend.

The study offers an explanation and flags it as probable, not established: the engine may be leaning more on what it memorized during training and less on the live web. It describes a "likely" algorithmic shift.

Worth being precise here: the study counts citations, not site rebuilds. It says nothing about what the affected sites did or didn't do over those seven weeks. But a simultaneous drop across four markets on a single day cannot come from the sites themselves — nobody rebuilds their site in Germany, Canada and the United States on the same morning.

seoClarity's own conclusion deserves quoting straight: this is not a permanent step down but volatility — "sharp moves in both directions, on OpenAI's timeline rather than the market's."

A model release swaps the instrument, not the site

Two clocks govern your AI visibility, and only one of them is yours.

The first is your site. It changes when you decide, at a pace you know. The second is the engine doing the measuring, whose balance between memory and live retrieval can move overnight, unannounced, at a scale measured in tens of percent — the episode above reaches 94%.

A model release is a move on the second clock. On July 24, Opus 5 became the default for Claude Max subscribers and the most capable model offered to Pro subscribers: overnight, Max subscribers' conversations were running on a different model. Does that switch change which sources get cited? Nobody can say yet. But if your numbers move over the next few weeks, you won't be able to say either.

That's where the expensive mistake of 2026 lives, and it isn't a technical one. It's reading a short-window swing as a result. We've all done it: the lovely three-week climb you put in the deck, the three-week slide you apologize for. Either way you're narrating something whose cause you don't know. It's a bet dressed as a metric — and sooner or later the swing arrives at the worst possible moment.

What actually got cheaper about Opus 5

A short detour through the release itself, because the headline number gets misread.

Opus 5 costs 5 dollars per million tokens you send it and 25 dollars per million it produces. That is the exact price of the previous version: the sticker price did not move. What falls is the cost of finishing a job, because a model that fails less often retries less often. Anthropic almost always compares its models on cost per task rather than per million tokens, and positions Opus 5 near its top-end model, Claude Fable 5, at roughly half the price.

Two caveats before you read that as a saving. Output costs five times what input costs, and Opus 5 writes longer answers than its predecessors — so the real bill can climb while the rate card stays flat. The full breakdown, including what it does worse and what breaks on migration, is in our dedicated piece: Claude Opus 5, what actually changed. These are the vendor's own measurements, published as charts, and no independent lab has reproduced them yet.

The consequence for you is indirect but real. When an assistant's work gets cheaper to complete, people hand it more of it — longer chains, deeper research, more pages read by a machine rather than a person. Your audience keeps drifting from the click toward the citation. It tells you nothing, though, about how those pages get picked.

How to measure when the instrument keeps moving

Four rules. None needs a specific tool — only the nerve to hold the line when someone wants an explanation.

1. Measure what you control first. The state of your pages — AI crawler access, structured data, how cleanly you answer — depends on no vendor. It's a stable score that moves only when you move it. That's your baseline, and the only thing you can responsibly commit to.

2. Watch for longer than an episode lasts. March–April took close to two months to resolve: on a six-week window you'd have mistaken the trough for a trend. Give it a full quarter before concluding anything.

3. Always compare against a basket of competitors measured over the same window. This is the most useful rule here. If three sites in your category drop the day you do, the cause isn't yours — you just saved a week of investigation. If you drop alone, the problem is genuinely yours and worth chasing.

4. Track more than one engine. This runs against instinct, because traffic concentration argues for the opposite: ChatGPT sends 92.4% of visits referred by AI chat platforms, per the Previsible report from July 2026 — a figure that excludes answers generated inside Google. That's exactly the point: a panel limited to ChatGPT puts your entire measurement in one vendor's hands, on a calendar that isn't yours. A second engine isn't there to bring traffic, it's there as a control. Our engine-by-engine breakdown covers where that traffic actually lands.

Together these get you out of day-to-day commentary. Citation tracking stops being a screenshot for the Monday meeting and becomes what it should be: an instrument whose margin of error you actually know.

What doesn't move

The reassuring part, and it holds up: none of this undermines the underlying work.

Pages crawlers can read, a clear self-contained answer up top, dated facts tied to a named source, complete structured data — those four held for ChatGPT in 2025, they hold for Opus 5 today, they'll hold for whatever ships next quarter. Our complete guide to GEO covers them with the method attached — GEO being simply the work that makes content quotable by answer engines, the way SEO makes it findable on Google.

Their real virtue, in the context of this article, is that they depend on nobody. When the measuring instrument shifts under you, the one thing still in your hands is the quality of what you expose. The rest runs on someone else's release calendar — and that calendar isn't yours.

The takeaway

• Opus 5's sticker price didn't drop (5 and 25 dollars per million tokens, same as before). What drops is the cost of a completed task — which means more of your pages read by machines.
• An engine's citations can fall 86% to 94% with no change on your side. Measured across five markets between February and April 2026, and recovering afterward.
• A model release swaps the instrument, not the site. Any reading shorter than the episode it spans confuses the two.
• Your only stable baseline is the state of your pages — the one thing neither OpenAI nor Anthropic can move on a Tuesday morning.

That baseline is exactly what the LightSpot audit measures: 46 criteria across 3 pages, free, with a score out of 100 and the 3 fixes worth doing first. It's the number that moves only when you move it — and the starting point our agents work from when applying the fixes. Citation tracking comes after, with the caveat attached: several engines, several weeks, and competitors alongside you, so that the day something moves, you know whether it came from you.