---
title: 'Google Trends for keyword research: five traps we hit in our own data'
description: >-
  We used Google Trends twice in one week to pick blog topics. Five traps showed
  up in our data, from a sinking anchor term to queries that mean something
  else.
author: elroot_labs
date: '2026-11-18'
categories:
  - digital-infrastructure
---
Google Trends is useful for keyword research when you treat it as a direction-finder and check every reading before you act on it. It tells you whether interest in a term is rising or falling relative to itself and to the other terms in the same comparison. It never tells you how many people searched, and the related-query lists mix real demand with noise. We used it twice in one week, on 14 and 16 November 2026, to choose posts for this blog, and five traps showed up in our own numbers. Each one comes with the check we now run.

## Trap 1: every comparison has its own 100

Trends scales each comparison so its highest weekly point is 100. The same term therefore gets a different number every time you change its neighbours. In our 16 November pull, "dpdp act" averaged 19 over the past year in India when compared with "ai automation" and "vibe coding", 16 next to "supabase" and "google ai mode", and 58 next to "llms.txt" and "schema markup". The week of 16 November 2025, when the DPDP Rules had just been published, scored 33 in one comparison and 100 in another.

None of those numbers is wrong. They are on different scales. The check: never copy a Trends figure into a spreadsheet without the other terms it was compared with. If you need one scale across many terms, keep one shared term in every comparison and divide by it, which is what our first study did.

## Trap 2: the anchor term can move under you

Our 14 November study put "robotics" in all 32 comparisons as the shared anchor, then measured each term's direction relative to robotics. Two days later we looked at robotics on its own, worldwide. It peaked at 100 in the week of 10 May 2026, stayed between 89 and 99 through June, then fell to 47 by late September and 25 to 29 in October and early November.

A term that held perfectly flat in absolute interest would have looked like it was climbing against an anchor that lost more than two thirds of its height. Several "up" readings in the first study need that caveat. The check: before you use an anchor, chart it alone for the same period and region. Pick one that is steady, or report directions in each term's own scale.

## Trap 3: the last week is always short

Weekly data includes the current week while it is still in progress. On 16 November, the final point for the week of 15 November was low for nearly every term we pulled in India: "dpdp act" 11 after 19 the week before, "ai automation" 22 after 28, "vibe coding" 8 after 12. That is one or two days of data, not a collapse.

The check: drop the final week before you compare recent weeks with earlier ones, or wait until it is complete. This one is easy to miss on a chart, because the line drops at exactly the point your eye goes to.

## Trap 4: related queries carry noise and other meanings

The rising and breakout lists are where most topic ideas come from, and they need the most filtering. In our pulls:

- Worldwide, the rising queries for "dpdp act" included "new york times" and "best laptop for work", both marked Breakout, neither related to Indian data law.
- In India, "website speed test" is mostly about internet connections. Its rising queries were "down detector website", "data speed test" and "how to test internet speed". We had planned a post on web page speed from that term and dropped it.
- Two Breakout queries under "ai voice agent" in India were quiz questions about one vendor's platform, which looks like learners searching for course answers rather than buyers.

The check: search the query yourself and read the first page of results before it goes on the list. If the results are about something else, so is the demand.

## Trap 5: the sample changes between downloads

Trends is computed from a sample of searches, and the sample changes. Medeiros and Pires [show that two samples differ](https://arxiv.org/abs/2104.03065) even with the same term, date range and location, and warn that this can produce conclusions that arise by chance. A 2025 paper by Djorno, Santillana and Yang describes [daily sampling variation](https://arxiv.org/abs/2504.07032) that causes discrepancies between historical downloads, and privacy thresholds that turn low search volumes into zeros.

We saw the small-number version of this directly. Worldwide, "ai voice agent", "deepfake detection" and "dpdp act" averaged 2, 1 and 1 next to robotics. At that size, a change from 1 to 2 doubles the reading, and a 0 can mean "too few searches to report" rather than none. The check: for any term below about 5, pull it again on another day, or compare it with smaller terms so it is not squashed against a large one. Do not report a percentage change on a term that small.

## What survived the checks

After those checks, the useful signals from our 16 November pull were plain enough to plan on:

- "google ai mode" in India rose from 24 in the week of 16 November 2025 to 87 in August 2026 and 74 in early November, with a Hindi query ("एआई मोड") marked Breakout. We wrote about it in [what Google AI Mode means for your website](/blog/google-ai-mode-website).
- "rich results" and "google rich results test" were the top rising queries under "schema markup" in India, which led to [our structured data audit](/blog/schema-markup-audit).
- "vibe coding" in India averaged 23 against 64 for "ai automation", with "vibe coding business impact" marked Breakout. That became [a post on taking an AI-built site to production](/blog/vibe-coding-to-production).

## A routine that fits in an hour

This is the routine we will use next time:

1. Choose an anchor and chart it alone first. Keep it in every comparison.
1. Use the past 12 months, weekly, and drop the final partial week.
1. Record every comparison with its full list of terms, region and date, so the numbers can be read later.
1. For each rising query you might write about, read the first page of real search results.
1. Re-pull anything below about 5 on another day before you trust its direction.
1. Treat Trends as one input. A topic still needs something only you can say about it, which for us means a test, a measurement or our own build history.

Trends is free and fast, and these checks take less time than writing one post on the wrong topic. For how AI search engines pick up the posts you do write, see [our earlier piece on AI search optimisation](/blog/ai-search-optimization-chatgpt).
