I track a commercial site's position on about a dozen money queries every Monday, two ways. A script calls a search API with the location set to the United States. And a browser agent, an AI that drives a real browser, runs the same queries in a private, signed-out window and reports what Google shows: positions, the top three results, whether an AI Overview appears and who it cites.
Two methods, so that one can catch the other's mistakes.
This Monday they disagreed about everything.
The script said my site had slipped from #1–2 to #4–6 on its main queries over two weeks, and a competitor now held #1 on all of them.
The browser check said my site was #1 on three of those queries and was the first source cited in Google's AI Overview. The same competitor had fallen from #1 to between #12 and #16. Its "changes since last week" line was a list of collapses.
One of these meant "you're losing", and the other meant "you're winning, and your rival just fell off a cliff". I had a page edit half-planned based on the first one.
Both were right
The agent's answer started, as the request asks it to, with the location Google displayed at the bottom of the results page. It wasn't the US. The browser was running on a connection in another country, the one the agent's machine actually sits in. In previous weeks it had been routed through a US connection. This week that route was gone, and nothing in the job noticed.
The data wasn't wrong. It was a faithful snapshot of what Google shows a searcher in that country. My buyers aren't searching from there.
A ranking isn't a property of the page. It's a property of the searcher.
I had been treating "position" as a fact about my page that a check measures, like page weight or response time. It isn't. Google ranks a page for a person: a country, often a city, a language, sometimes their history. There is no position without a location, and a check that doesn't record its location hasn't measured anything you can compare.
Not just country: city too
Once I started paying attention to location, I added a second view to the scripted check: the same queries pinned to the main city in the market, alongside the generic US results. On most queries they agree. On some they don't:
| Query | US-wide | Pinned to the city |
|---|---|---|
| Main commercial query | #2 | #4 |
| Secondary query | #6 | #4 |
| Local rentals query | #7 | #9 |
Neither column is "the real ranking". People planning a trip search from home, all over the country. People already in the city search from there, and Google weights local businesses more for them. Both audiences matter, and they see different pages.
What changed
Every result records its location. The browser check now has to report Google's displayed location on the first line, and the review step that reads its answer checks that line before anything else. This week that review is what caught the problem. It flagged the answer as "wrong location, don't compare with last week" before I acted on it.
A run from a different location is not a data point. It doesn't go into the history, and it doesn't count toward "dropped two weeks in a row" alerts. Comparing this week's foreign-location run with last week's US run would have produced exactly the dramatic swings it showed, out of nothing.
Pinned locations beat "wherever the checker is". The weekly job moved entirely to the scripted check, which sets the location explicitly, once US-wide and once pinned to the city, with a separate history file for each so neither view's week-on-week comparison is polluted by the other. The browser agent still does the things only a browser can, like reading AI Overviews in full. It no longer owns the number.
Search Console isn't a tiebreaker for this. Its "average position" blends every location, device and query variant into one figure. Useful for trends, useless for "where do I rank for a buyer in this market today". I'd already been burned by a page that ranked #6 and #20 for the same query and averaged #13.
The general version
Any measurement that depends on who is asking, such as rankings, prices, ad auctions or what an AI answer says, needs the "who" recorded next to the number, and needs it held constant between runs you intend to compare.
My check didn't fail. It answered a question I hadn't asked, precisely and confidently, and it looked exactly like the answer to the one I had asked.
![Quadruped Robot Software — Ch.10: Quadruped AI and Physical AI [KR]](https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Ffiibwtdu8wij6gnnuoyc.png)

