In the last week of March, Google Trends put "claude code" at 100, its peak for the year. By the end of October it was at 37, down 49% in the latest quarter. Had I looked only at that line, I would have stopped writing about Claude Code.
Search Console says the opposite. My daily impressions for queries containing "claude" roughly doubled between the first half of July and the second half of October. Demand shrank, and my share of what was left grew.
Neither number is the decision. The decision comes from reading both together, and then asking a third source whether that traffic was ever worth anything.
That join used to cost me an afternoon of exports and a spreadsheet. It now costs one prompt. Every number below comes from runs on my own site on 3 and 4 November 2026, with Trendzeist 0.6.
What each source can and cannot tell you
Each source answers one question well and fails at the other two.
| Source | Answers | Can't answer |
|---|---|---|
| Google Trends | Is interest in this topic growing or shrinking, and where? | How many people searched. Every number is a 0 to 100 index relative to the peak in your window. |
| Search Console | Does Google already show my site for this, at what position, with what CTR? | Anything about queries you don't rank for. A large share of low-volume queries also arrives anonymized. |
| Google Analytics | Did visitors from that page stay, read, and move toward a service or a form? | Why they searched in the first place. |
Trends without Search Console hands you topics you have no chance of ranking for. Search Console without Trends has you polishing pages for demand that is already leaving. And both without Analytics will happily point you at traffic that never buys anything.
The setup: three MCP servers, all read-only
# Google Trends, Autocomplete, News and Wikipedia. No API key, no account.
claude mcp add trendzeist -- uvx trendzeist-mcp
# Search Console: a wizard that signs you in and writes the config
npx -y suganthan-gsc-mcp setup
# Google Analytics 4: Google's own server
claude mcp add analytics-mcp --scope user \
-e "GOOGLE_APPLICATION_CREDENTIALS=PATH_TO_CREDENTIALS_JSON" \
-e "GOOGLE_CLOUD_PROJECT=YOUR_PROJECT_ID" \
-- pipx run analytics-mcp
Trendzeist is mine: a free, MIT-licensed server I built because the Google Trends libraries I tried were abandoned, needed a headless browser, or got rate-limited within the hour. The Search Console server is Suganthan Mohanadasan's, and his setup guide covers the one-time OAuth client step better than I would. The GA4 server is Google's own and needs Application Default Credentials with the analytics.readonly scope. Claude Desktop and Cursor take the same commands in their JSON config.
Keep all three read-only. An opportunity analysis never needs to write anything, and an agent that can't change anything can't break anything.
The prompt
Analyze my website's organic search opportunities for the next 30 days.
Use Trendzeist to identify rising, breakout, and rapidly accelerating
topics related to my existing content.
Use Google Search Console to determine whether my site already receives
impressions for those topics or closely related queries.
Use Google Analytics to determine whether traffic from those topics/pages
produces meaningful engagement or conversions.
Prioritize topics where:
- search interest is increasing,
- my site already has some topical relevance,
- there is an achievable ranking opportunity,
- and similar traffic has demonstrated business value.
For each opportunity, recommend:
- CREATE new content
- UPDATE existing content
- EXPAND an existing topic cluster
- IGNORE
Return the top 10 opportunities with evidence from all three sources.
When a source has no data for a topic, write "no data". Do not fill the gap.
Most of the weight sits in two lines. "Evidence from all three sources" forces the model to show its work, and the last line stops the most common failure I see in these analyses: a confident row built from one source and two guesses.
Step 1: Trendzeist finds the moving topics
Claude's first call looked like this:
discover_topics(
seed_keywords=["claude code", "n8n"],
timeframe="today 3-m",
site_queries=[...24 queries pulled from Search Console...],
)
site_queries is the parameter most people skip. Claude pulls the queries my site already ranks for from Search Console and passes them along, so every topic comes back tagged as covered, partial or gap.
Google's raw rising list is the first lesson. Its top breakouts under "claude code" and "n8n" were "best time to visit maldives" (+7,150%), "how to bake a cake" (+6,250%) and "top tourist attractions in paris". The rising list for a broad term picks up whatever a few people typed next to it, and for AI tools that includes the prompts people paste into the search box. Trendzeist moves those into a filtered_out list, because they share no word with the seed or its top searches, and its note suggests adding a category. Spam goes to a separate suspect list: here, a self-promoting domain and two queries carrying the same person's name. Both checks work on words, so treat them as a first pass, not a verdict.
Rerunning with category=5 (Computers & Electronics) cleared the junk. The interesting part was in filtered_out:
| Rising query (seed: "claude code", category 5) | Growth, last 90 days | Where Trendzeist put it |
|---|---|---|
| deepseek harness | +2,500% | filtered_out |
| omniroute | +1,000% | filtered_out |
| herdr | +600% | filtered_out |
| agent router | +250% | filtered_out |
| omniroute github | +650% | topics |
| claude code harness | +100% | topics |
| claude mem | +60% | topics |
There's a cluster in there: people want to run Claude Code against other models, through routers and harnesses. Four of its queries share no word with "claude code", so the word check set them aside. Any automatic filter makes that trade, which is why you read filtered_out before you ignore it. Before calling the cluster an opportunity, Claude checked the 12-month curves with interest_over_time. "claude code harness" peaked at 13 in late March and sits at 6. "claude code router" fell 64% in the last quarter. On the same chart, "claude code skills" peaked at 100 in May and sits at 20 now.
A 2,500% rise from almost nothing is still almost nothing. Trendzeist says as much in its own output, where every rising-only item carries "popularity unknown" in the score breakdown. That's the tool telling you to look at the curve before you write.
Step 2: Search Console shows where you already stand
Claude pulled 90 days of queries, 5 August to 2 November. The rows that mattered:
| Query | Where it stood |
|---|---|
| n8n consultant | Several hundred impressions on page four, no clicks |
| competitor checker | A couple hundred impressions on page seven, no clicks |
| rag consulting | A couple hundred impressions on page three, no clicks |
| claude code zsh | Lower half of page one, the odd click |
| claude code wordpress | Between the bottom of page one and the top of page two |
| zsh claude | Page one, the odd click |
| n8n vs openclaw | Page two |
I've rounded my own figures into bands here. On your site, the prompt returns the exact impressions, clicks and positions.
Page totals fill in the rest. My macOS setup tutorial and my Claude for WordPress guide both rank near the bottom of page one and get clicked. The n8n service page shows up for buyer queries around page six, and nobody clicks.
Coverage matching handles near-variants: "claude code mac" counts as covered by "claude code macos", and every verdict names the site query behind it in matched. Read that column before you drop a topic as covered, because a match on words can't tell whether your page actually answers the query. And feed it the messy query variants Search Console shows you, not a tidy list.
Step 3: Analytics shows which traffic is worth anything
This step reshuffled the whole ranking. Organic sessions over the same 90 days:
| Landing page | Organic sessions | Engagement rate | Avg. session |
|---|---|---|---|
| Claude for WordPress development | 113 | 60% | 2 min 58 s |
| Claude Code macOS setup | 87 | 57% | 3 min 10 s |
| Claude Projects for consultants | 73 | 53% | 4 min 18 s |
| Claude Code on Windows with WSL | 32 | 72% | 3 min 16 s |
Engagement looks healthy. The business column doesn't. The key events column was zero for every landing page, so Claude used the proxy I gave it: organic sessions that started on a blog post and later opened a service page or the About page. In 90 days, that added up to six page views.
The tutorials still matter, since they're the reason the site has impressions at all. But on this site a tutorial's job is visibility, and the pages that pay the bills are the service pages, which are exactly the ones Search Console shows on pages three to six. Skip this step and the analysis will keep recommending more tutorials, because tutorials are what traffic numbers reward.
The ranked list
| # | Topic | Trends | Search Console | Analytics | Call |
|---|---|---|---|---|---|
| 1 | n8n consultant | "n8n" down 49% in the last quarter | page four, no clicks | no organic sessions on the page | UPDATE the service page |
| 2 | RAG consulting | "rag chatbot" down 61% | three variants on page three, no clicks | no organic sessions on the page | UPDATE the service page |
| 3 | zsh setup queries | "claude code" down 49%, my share up | six zsh variants on page one, a few clicks | 87 sessions, 57% engaged | UPDATE the snippet |
| 4 | Claude Code + WordPress | no trend signal | page one to page two across variants | best landing page: 113 sessions | EXPAND the guide |
| 5 | Claude Code routers and harnesses | rising from near zero, now flat | no data | no data | CREATE, low priority |
| 6 | Claude Code skills | 100 in May, 20 now | existing post: a trickle, mostly in Portuguese | negligible | IGNORE as new content |
| 7 | n8n vs OpenClaw | "openclaw" down 64% | page one, low CTR | 13 sessions, 5 min avg. | UPDATE the title only |
| 8 | competitor checker | no data | hundreds of impressions, page seven and beyond | none | IGNORE |
| 9 | Maldives, cakes, Paris | raw breakouts, set aside as off-topic | no data | no data | IGNORE |
Rows 1 and 2 rank first despite falling trends because someone searching "n8n consultant" is shopping. The fix for row 1 is to put "n8n consultant" in the page's title and H1, and to merge a competing blog post into it with a 301, because that post was collecting impressions for the same query and almost no clicks.
Row 3 was the cheapest fix on the list. Six zsh variants ("claude code zsh", "add claude to zsh", "zsh status line"...) sit on page one, and the page's meta description didn't mention zsh. Adding the word is a one-line change.
Row 7 would fool me without Trends. The page ranks well, and the topic is collapsing, from 100 in March to 2 now. A new title is fifteen minutes of work. A rewrite would be a week spent on a topic whose readers are leaving.
The prompt asked for ten. I kept nine, because the tenth candidate had evidence from a single source. A list that admits it ran out of evidence beats one that fills the quota.
The near-miss version
My second prompt narrows the same analysis to pages that are already close:
Find trending SEO opportunities where my website is already close to
ranking well.
Use Trendzeist to identify topics and queries with rising or breakout
search interest. Cross-reference those queries with Google Search Console
and find: average position between 5 and 30, meaningful impressions,
low or below-potential CTR, pages already ranking for the topic.
Then use Google Analytics to determine whether visitors landing on those
pages show meaningful engagement or conversion behavior.
Give me the 10 highest-priority optimization opportunities and explain
exactly what should be changed.
On my site it returns rows 3, 4 and 7, plus the RAG page, which falls to the bottom because no organic visitor has landed on it yet. This is the version I run every month. The full one is quarterly.
Where this goes wrong
- Trends is relative, always. A topic at 6 can grow 100% and still be smaller than anything you already rank for. Never accept a growth percentage without the curve.
- Broad seeds return noise. Add a category, or seed with the phrase your buyer actually types, which is the same reason I start keyword research from customer problems instead of volume.
- Search Console hides queries. Anonymized queries never show at query level, so page totals are always bigger than the rows you can see.
- Positions are blended averages across devices and countries. An average of 35 can be page one in one market and nowhere in the others.
- Without key events, Analytics measures attention. Set up a key event for the contact form and the booking link before you run this, or tell the model which proxy to use.
- The model will rank whatever you give it. Read the evidence columns before the recommendation column.
Run it on your own site
Install the three servers, paste the prompt, swap my seeds for the two phrases your customers use, and read the evidence before the verdicts. The first run takes around ten minutes. Most of that is Trendzeist waiting at least two seconds between requests to Google on purpose, which is why it keeps working after quick scripts get rate-limited. If you'd rather run this on a schedule than on demand, it's the first half of the n8n SEO pipeline I built for a client. The next post in this series asks whether your visibility is keeping pace with demand.
If you'd rather get the analysis with judgment attached, this exact workflow is my Search Opportunity Audit: I run it on your site, check every row by hand, and give you a 30-day plan in priority order. To see whether that fits, book a free 30-minute Intro Call.
Which of your best-ranking pages is ranking for a topic that is already leaving?
