The phrase "MCP gateway" is used as if it named a settled category. Eighteen paid measurements
around that object say something less settled: 91% of the demand sits in one phrase written three
different ways, and the rest fragments into thirteen phrases that each appear once.
Short version, from the data itself: eighteen phrases, 5,300 searches a month between them, of
which 4,800 are mcp gateway, mcp gateways and mcp-gateway — three spellings that return
byte-identical figures, right down to the cost-per-click. The remaining 500 searches are spread
across fourteen clusters, thirteen of which are a single phrase. Anyone deciding what to document,
what to build, or where to spend the next month of work should see that shape before they start.
Three spellings, one number, to the cent
The first thing the measurement produced was not a volume but a duplicate:
| Phrase | Monthly searches | Cost per click |
|---|---|---|
mcp gateway |
2,400 | $49.58 |
mcp gateways |
2,400 | $49.58 |
mcp-gateway |
2,400 | $49.58 |
Byte-identical across volume, bid range and the twelve-month series behind them. The advertising
endpoint treats the singular, the plural and the hyphenated form as one term, and it is right to:
no reader distinguishes them, and no search engine does either. The practical consequences are
small but easy to get wrong — you cannot add those three numbers together to claim a 7,200/month
category, and you should not build three pages to cover them. One page that uses all three
spellings in its prose and its title is the whole job.
The expensive part is worth a second look: a $49.58 cost-per-click on a developer-infrastructure
phrase means advertisers are paying real money to intercept someone at the moment they ask what
this thing is. That is a strong hint about the intent behind the head term — the searcher is not
shopping for a protocol, they are trying to place an unfamiliar object.
The rest is people arriving from their own stack
Take the head term away and the tail is not a topic. It is a list of the tools people already run:
-
microsoft mcp gateway170 -
what is an mcp gateway140 -
best mcp gateway70 ·model gateway70 -
mcp server gateway40 -
okta mcp gateway30 -
api gateway for ai20 ·github mcp gateway20 ·mcp gateway vs mcp server20 ·mcp vs api gateway20 -
agent gateway open source10 ·docker mcp gateway compose10 ·envoy ai gateway mcp10 ·amazon api gateway adds mcp proxy support10 ·ai gateway vs mcp gateway10 ·best open-source mcp gateway10
Okta, GitHub, Docker, Envoy, Amazon API Gateway — those are not four competitors and a duopoly.
They are five different front doors into the same question, each opened by someone who already has
one of those tools on their laptop. A vendor-qualified search is usually an integration question
wearing a product query's clothes, and it is why the long tail here looks wide but shallow.
The only cluster that survives
Clustering the eighteen phrases by how much their search results pages overlap — merge when two
phrases share five or more results, form a cluster at three, ignore domains that appear in more
than half the pages (here: github.com, linkedin.com, reddit.com, youtube.com and one commercial
docs site) — produced fourteen clusters. Thirteen were single phrases.
The one that was not:
-
what is an mcp gateway140 -
best mcp gateways70 ·best open-source mcp gateway10 -
mcp gateway vs mcp server20 ·ai gateway vs mcp gateway10
Roughly 250 searches a month of definition, comparison and selection intent, held together by
the only overlapping results in the whole set. Two pairs merged directly: the two "best" phrases
shared eleven results, and mcp gateway vs mcp server shared five with what is an mcp gateway.
When the results for "what is this", "which is best" and "how does it differ from that" are the same
pages, the searcher is doing one thing: deciding whether this object belongs in their architecture.
That is the shape of the demand, and it says where the writing goes. Not eighteen pages. One
explanation that answers the definition, the comparison and the selection in the same text — which
is what the reference page on MCP gateways is
built as, and why its title carries all three spellings rather than one.
Why a gateway attracts definition questions
A comparison cluster forms when an object is close enough to something familiar that people can only
describe it by subtraction. "Gateway" collides with "proxy" and "server" for a reason: those are
different layers with different jobs, and the difference is not visible from a config file.
A server answers questions about tools — which exist, what arguments they take, what a call returns.
A proxy moves bytes and can apply a limit to a path, but it does not parse a tool call, so it cannot
meter one. A gateway is the layer that answers questions about callers: which key is this, how many
calls it has made in this window, what this team has spent this month, and which record proves the
call happened. None of those four questions belong inside a tool implementation, which is exactly why
the object exists and why it is hard to name by analogy.
What the shape implies for anyone shipping a developer tool
Four things follow from the numbers, and none of them require a large budget to act on.
Pick one spelling and use the others in prose. The plural and the hyphen are not separate
markets, they are the same term. Covering them is a sentence in the first paragraph, not a second
URL — and a second URL would compete with the first for the same query.
Vendor-qualified long-tail belongs in sections. docker mcp gateway compose at ten searches a
month is a real question from a real person, and it is a paragraph, not a page. Twenty such phrases
do not add up to a topic; they add up to a configuration section.
Definition and comparison intent is the cluster worth owning. It is the only place where the
result pages agree with each other, which means it is the only place where one good page can be the
answer rather than one of forty.
Read the CPC next to the volume. A high bid on an informational phrase is evidence that the
people clicking it convert somewhere downstream. It is a better signal of commercial value than the
search count.
Method, and what this cannot tell you
The measurement is eighteen phrases queried against the Google Ads volume endpoint for the United
States with a twelve-month window, on 2026-09-28, through a paid account. The clustering used depth
20 search results per phrase on the same day; the merge and cluster thresholds were five and three
shared results, with ubiquitous domains excluded at a 50% share. Total spend on this piece of work
was under a tenth of a dollar, which is the other reason to do it before writing anything.
Three honest limits. The endpoint merges near-variants, so the three spellings above cannot be added
together but the tail's small numbers are still rounded to the nearest ten. A single measurement in
one country at one moment is a snapshot, not a trend; the same query a month later will move. And
volume is not traffic — a phrase can carry thousands of monthly searches and still send four visitors
to a page that ranks tenth.
The raw numbers are in the table above so the reading can be checked rather than taken on trust, and
the phrase set is small enough to re-run in an afternoon.
Disclosure: this measurement was run by SmartGate, which builds an MCP gateway — the object being
measured. The keyword data was bought with our own budget and is quoted in full above; the
interpretation is ours and is arguable, which is why the method is included alongside it.
Top comments (0)