The short answer
Take each client's GA4 conversion rate over the same 90 days. Leave out clients below a session floor. Group the rest by vertical and by what they count as a conversion. Then report each group's median and its 25th and 75th percentiles. A client's place is its quarter of that range, shown without naming anyone else.
- Each client gets one vote: rank client rates, never pool them into one figure
- Write the floor and the groups down before you look at a single rate
- Show clients the band, never the best or the worst client
Why your own clients beat published averages
A published benchmark answers a question about someone else's accounts. You can't see what counted as a conversion, which traffic went in, or which months it covered. Your own roster is different: you can open any number in it and check.
Definitions cause most of the damage. A standard GA4 property can mark up to 30 events as key events (as of November 2026), and no two account managers star the same ones. Your decking client may count quote requests and nothing else, while the kitchen fitter beside it also counts brochure downloads, phone taps and newsletter sign-ups. Line those two up and the fitter looks like the better seller, for reasons that sit in its GA4 Admin screen rather than on its website. That is why the method below groups clients by what they count before it ranks anyone. For how GA4 itself tallies key events and which rate DuckDive's tools return, see the GA4 integration page.
AgencyAnalytics, which publishes benchmarks, uses medians for the same reason: its post says it chose median values "because they provide a more accurate representation of industry norms", and that medians "eliminate the impact of outliers" (post dated February 12, 2025). That logic works even better on data you can audit line by line.
Your roster also shares one window, to the day, and you can open the landing pages behind any odd figure.
Medians and quartiles, not means
Rates across a client list are lopsided. Most clients bunch together, and one or two sit far out: a broken tag, a duplicate key event, a niche where every visitor arrives ready to buy. The mean chases those clients. The median ignores them. NIST's statistics handbook says it plainly: when the tails hold extreme values, "the median provides a better estimate of location than does the mean".
The middle half matters as much as the middle point. The 25th and 75th percentiles mark the edges of normal. The gap between them is the interquartile range, which NIST describes as less affected by extremes than the standard deviation.
One vote per client
Rank client rates. Don't pool them. A pooled rate, all conversions over all sessions, answers "how does our whole book convert", and your biggest client decides it. A benchmark asks what a normal client looks like, so each client counts once.
Your agent and your spreadsheet may disagree
There is more than one way to compute a percentile. NIST's page on percentiles labels the nine methods Hyndman and Fan compared R1 to R9. The handbook's default, R6, and the method Excel uses, R7, both interpolate between two neighbouring clients. They differ only in where the cut falls: at p(N+1) for R6, at 1 + p(N - 1) for R7. In the example below they differ by up to 0.15 points. Pick one method, write it in your notes, and ask your agent to say which it used.
- Conversion rate, for this method
- One client's conversions divided by its sessions over one window. Conversions are the GA4 key events DuckDive's tools report.
- Median
- The middle client once rates are sorted. Half sit above it, half below.
- 25th and 75th percentiles (Q1 and Q3)
- The rates a quarter and three quarters of the way up the sorted list.
- Interquartile range (IQR)
- Q3 minus Q1: the width of the normal band.
- Sample floor
- The fewest sessions a client needs in the window before its rate counts.
- Fences
- Q1 minus 1.5 IQR, and Q3 plus 1.5 IQR. A rate outside them is a mild outlier in NIST's terms; outside 3 IQR, an extreme one.
Ask your agentFor every client with at least 1,000 sessions in the last 90 days, give me GA4 conversion rate, then the median and the 25th and 75th percentiles.
Tools it calls list_websites get_ga4_overview
A minimum sample before a client counts
A small site's rate jumps for reasons that have nothing to do with its website. Two extra quote requests in a quiet month can lift it from the bottom quarter to the top. Leave that client in, and it drags the quartiles around with it.
| Sessions in the window | One more conversion adds |
|---|---|
| 200 | 0.50 points |
| 500 | 0.20 points |
| 1,000 | 0.10 points |
| 5,000 | 0.02 points |
A floor of 1,000 sessions over 90 days keeps that one-conversion swing to about a tenth of a point. It's a working rule, not a law of statistics. For a low-traffic niche, use a longer window instead of a lower floor; through DuckDive the next step up from 90 days is 6 months.
Clients under the floor don't vanish. List them under the benchmark with their sessions and conversions and judge them on their own trend.
Ask your agentWhich clients missed the 1,000-session floor over the last 90 days? Show their sessions and conversions, then tell me which of them would clear it on the last 6 months.
Tools it calls list_websites get_ga4_overview
Set the rules first
Decide the floor, the window and the groups before you see any rates. Otherwise you will nudge the floor to rescue a favourite client.
Grouping clients so comparisons are fair
One benchmark for the whole roster puts a personal injury firm next to a furniture shop. Group first, then compute. Two cuts do most of the work.
Cut one: vertical
Clients in one trade share buyer intent and sales cycle. Put the vertical at the start of each website's name in DuckDive, such as "Home improvement: Hatchwell Decking". A search on the vertical then returns the whole group.
Ask your agentGroup those sites by the vertical at the start of their DuckDive website name and give me the median for each group.
Tools it calls list_websites get_ga4_overview
Cut two: what counts as a conversion
Inside a vertical, split again by the action being counted. A quote form, a click on a phone number and a showroom booking convert at very different rates on the same traffic. DuckDive's tools report conversions without event names, so this label comes from your setup notes or a look at each property's key events in GA4 Admin. The Google Analytics 4 integration page covers what the tools read and how key events feed them.
Traffic mix, noted not grouped
A client living on paid search converts differently from one living on organic. Note each client's paid share beside its rate with get_ga4_channels rather than adding a third cut. SEO-led shops can benchmark organic landing pages only, alongside the SEO agencies triage routine.
How small is too small
Size sets what a group can support, and the same three tiers apply everywhere on this page:
- Eight or more counted clients: median plus the middle-half band, for clients and leadership alike.
- Five to seven: clients see the median alone. The band stays internal, because with so few members each quartile sits right beside one recognisable rate.
- Under five: no benchmark at all. Merge the group into its nearest neighbour, kitchens with bathrooms for instance, or compare each client with its own history.
The dental groups page runs the same idea across one group's own practices.
Worked example: benchmarks for eighteen home improvement clients
A strategist is prepping the monthly review. Every site in the home improvement group starts with the vertical in its DuckDive name, so the agent can find the group on its own.
Home improvement group, last 90 days. Apply the 1,000-session floor, give me the median and quartiles, and flag anyone outside the fences.
list_websitessearched for "Home improvement" and found 18 client sitesget_ga4_overviewonce per site, last 90 days: sessions, conversions and rate
15 of the 18 sites cleared 1,000 sessions. The other three had 410, 760 and 930 sessions, so I listed them apart and left them out of the ranks. Each rate is conversions over sessions. Percentiles use Excel's method (R7 in NIST's list).
| Measure | All 15 sites | Without outlier |
|---|---|---|
| Lowest | 1.10% | 1.10% |
| 25th percentile | 2.30% | 2.25% |
| Median | 3.00% | 2.90% |
| 75th percentile | 3.75% | 3.53% |
| Highest | 9.80% | 4.60% |
| Mean, for contrast | 3.34% | 2.88% |
One site, a gutter guard installer, converts at 9.80%. That is past the outer fence: 3.75% plus three times the 1.45-point IQR is 8.10%. Its conversions leap on a single day mid-window while sessions hold steady, which points at tracking rather than customers. I left it out, so the benchmark is the right-hand column: 14 sites, median 2.90%, middle half 2.25% to 3.53%.
Dropping it moves the median by a tenth of a point and the mean by almost half a point. Report the median and the band, and leave the mean off the slide.
Look at what the median did when the odd site left: almost nothing. The quartiles are touchier about method than about outliers. On the same 14 sites, NIST's default R6 would put them at 2.13% and 3.68%, against 2.25% and 3.53% from R7. Neither is wrong; keep one from month to month. Before excluding a site for good, have whoever manages its property confirm in GA4 Admin which key event was added; the GA4 integration page shows what to look for.
Ask your agentUsing those quartiles, list any client above Q3 plus 1.5 times the IQR or below Q1 minus 1.5 times the IQR. For each one, show daily conversions for the last 90 days so I can spot a key event that changed.
Tools it calls get_ga4_overview
Get your roster's quartiles from one prompt
Link each client's GA4 property once, and your agent can rebuild every group's median and band whenever a review comes round.
Free for 2 sites · No credit card
Worked example: one client's place in its group
Hatchwell Decking's owner says leads feel slow. The strategist checks its place first, then what drags it there, in one question asked right after the group run above.
Where does Hatchwell Decking sit in its group, and which landing page pulls it down most?
list_websitesfound Hatchwell Decking in the home improvement groupget_ga4_overviewlast 90 days: 6,420 sessions and 122 conversionsget_ga4_landing_pageslast 90 days: top landing pages by sessions, with conversions
Over the last 90 days Hatchwell converts at 1.90%, below the group's 25th percentile of 2.25%. It sits in the bottom quarter, third from the bottom of the 14 counted sites.
| Landing page | Sessions | Rate |
|---|---|---|
| /composite-decking-cost | 2,310 | 0.39% |
| / (home) | 1,450 | 2.14% |
| /gallery | 820 | 1.34% |
| /free-quote | 640 | 9.06% |
| /pergolas | 560 | 1.61% |
| All other pages | 640 | 0.63% |
The decking cost guide brings in 36% of all sessions and converts at 0.39%. Without it, the rest of the site converts at 2.75% (113 conversions from 4,110 sessions). That would put Hatchwell in the second quarter, between 2.25% and the 2.90% median. The quote page itself is healthy at 9.06%.
So the gap comes from a research page doing what research pages do, not a broken funnel. Test a quote prompt on the cost guide before anyone talks about a rebuild.
One question remains before the owner hears any of this: is the cost guide's poor rate new, or has it always been that way? A page that slid from 1% is a fault to chase. One that has sat near 0.4% for a year is simply a research page, and the fix is a better path to the quote form.
Ask your agentCompare Hatchwell Decking's /composite-decking-cost page on the last 90 days with the last 12 months: sessions, conversions and rate for each. Has it always converted this low?
Tools it calls get_ga4_landing_pages
The rank tells you where a client sits. The landing pages tell you why. When search traffic is part of the story, the guide to pages that rank but don't convert takes the next step with Search Console data.
Showing a client its position without naming others
Clients like knowing where they stand. They would hate finding their own numbers on another client's slide. So the client version follows hard rules.
- Show positions, not people. The median and quartiles can equal a real client's rate. With an odd count the median always does, and the default R6 method often lands a quartile on one too. That is acceptable because the figure carries no name and sits in the middle of the pack, among several clients with similar numbers.
- Never show the lowest or highest rate. Each is one client, and a standout is the client a reader is most likely to guess: the biggest advertiser in the group, or the newest site.
- Watch small groups. With five to seven counted clients, show the median alone, because each quartile sits right beside one rate. Under five, show no benchmark at all.
- Describe the group, not its members. "14 home improvement sites we manage" is fine. "Three deck builders in your county" points at competitors.
- Round to one decimal. Extra digits make a rate easier to trace to one client.
| Item | Client review | Agency leadership |
|---|---|---|
| Group median | Yes, rounded | Yes |
| 25th to 75th band | Yes, as "the middle half" | Yes |
| Client's own rate and quarter | Partly, their own only | Yes, every client |
| Lowest and highest rates | No, each is one client | Yes, with names |
| Other clients' names | No | Yes |
| Clients below the floor | Partly, a count only | Yes, with sessions |
| Groups of five to seven | Partly, median only | Yes, full range |
| Groups under five | No benchmark shown | No, each client's own history |
Ask your agentWrite Fenmoor Kitchens a three-sentence note on where its 90-day conversion rate sits among our home improvement clients. Use the group median and the middle-half band only. No other client's name or rate, and no highest or lowest.
Tools it calls list_websites get_ga4_overview
"Over the last 90 days, Fenmoor Kitchens turned 4.2% of website visits into enquiries (163 from 3,880). Across 14 comparable home improvement sites we manage, the middle rate was 2.9%, and the middle half ran from 2.3% to 3.5%. That puts you in the top quarter, so we will protect what is working before testing anything new."
Explaining benchmarks to clients and bosses
To a client
Start with their own trend, then their position. "The middle half of comparable sites" lands better than "interquartile range". Say what counts as a conversion for them before any comparison. If they sit low, bring the landing page finding, so the meeting ends on a fix.
To your boss or agency lead
One table per vertical: median, band, how many clients counted, and who sits below Q1. Put the floor, window and percentile method in a footnote, so the number survives the first challenge. For quarterly review prep across services, the full-service agencies page folds this into a wider QBR routine.
When someone waves an industry average at you
Don't argue that the published figure is wrong. Say instead: "Ours uses our clients' definitions, one window for everyone, and we can open any number in it." That answer wins because it can be checked.
Ask your agentRerun the home improvement quartiles on the last 6 months and tell me which clients changed quarter compared with the 90-day run.
Tools it calls list_websites get_ga4_overview
A client that hops between quarters from one window to the next is noise. One that stays in the bottom quarter on both is a real conversation.
Before a benchmark goes in a deck
- Same window. Every client's rate covers the same fixed range.
- Floor applied. Clients under it are listed apart, with their counts.
- Group size checked. Eight or more for the band, five to seven for the median alone, under five for nothing.
- Conversion type matched. Every client in a group counts the same kind of action.
- Fences checked. Every rate past them is traced. A tracking change keeps it out of the ranks; a real result stays in, with a note.
- Method named. The percentile method and floor sit in the footnote.
- No names leak. The client version shows the band and their own rate only.
- Math spot-checked. Two clients' rates redone by hand from sessions and conversions, as the guide to checking your agent's numbers describes.
Loading your client roster into DuckDive
DuckDive is where your agent gets the numbers. You sign in with Google, link each client site's GA4 property, and DuckDive imports the history and stores it. Your agent then reads each site's sessions, conversions and landing pages through DuckDive's MCP server and does the sorting and percentile math in the conversation.
There is no benchmark button. The agent calls get_ga4_overview once per client from stored data, and you can ask it to list every figure it used. The DuckDive for agencies page shows the rest of what a roster can answer.
-
Sign in with Google
"Continue with Google" creates your DuckDive account with no form. Check one client's numbers against GA4 before adding the rest.
-
Add each client site, vertical first in the name
Agency accounts hold several clients, each with its own websites. Name each site "Vertical: Business". An agent can do the adding for you if you are an admin and kept "Also allow actions" ticked when connecting; the guide to adding dozens of websites with an agent covers a big roster.
Your agent can add the rosterSet up DuckDive for my website. Follow https://analytics.duckdive.ai/
agents.md -
Link every client's GA4 property
DuckDive lists the GA4 properties your Google login can reach. If a client's property is missing, someone with admin rights on it adds your Google account as a user. A property belongs to one DuckDive website at a time.
-
Let the history import
Linking starts the import. It usually takes a few minutes, longer for a long history. Ask your agent to run
get_sync_status, which shows each source as ready, importing or failing, before you trust a benchmark built on it. -
Connect your agent and run the first prompt
In your AI app, add
https://api.duckdive.ai/mcpas a custom connector or remote MCP server, then select Continue with Google on the DuckDive page. ChatGPT needs a Business or Enterprise plan for this. Claude Code takes a one-line terminal command and Cursor an install link instead; the Claude Code and Cursor setup pages give both, and the Claude setup page covers the web and desktop apps. If the setting is not where you expect, search the app's settings for "Connectors", "Apps" or "MCP", since vendors rename these menus often.
What this benchmark cannot tell you
- Anything outside your roster. Your agent sees only your account's sites. DuckDive supplies no industry figures.
- Custom date ranges. The tools take fixed windows: the last 7, 14, 30 or 90 days, 6 months or 12 months. Calendar quarters need GA4 itself for now.
- Key event names, devices or regions. The GA4 tools cover dates, channels or sources, and landing pages. Your conversion type labels come from your own notes.
- Unchecked math. DuckDive supplies the numbers; the agent does the math. Ask for the sorted rates.
- Campaign-level ad benchmarks. Google Ads data comes in as an account summary, so you can compare cost per conversion across clients, not across campaigns.
- Branded client reports. DuckDive does not build white-label PDFs. If you need those, the AgencyAnalytics comparison explains where that tool fits better.
- Why a client converts poorly. A low rank is a question for landing pages, offers and sales follow-up.
And if you have six clients across six trades, skip all of this. Compare each client with its own past instead.
What strategists ask before trusting the numbers
Why not just quote a published industry benchmark?
Because you can't check what it counted. Published figures mix definitions, traffic types and windows you never see. Your roster uses definitions you can inspect and one window for all. Mention a published figure as background if asked, but decide on the internal median.
How many clients do I need before quartiles mean anything?
Eight clients that clear the floor, within one vertical and conversion type, is a workable minimum for quartiles. With five to seven, show clients the median only. Under five, show no benchmark: merge the group with its nearest neighbour or compare each client with its own history.
Should I use the median of client rates or one combined rate?
For a benchmark, the median of client rates. Each client counts once, so one large account can't set the standard for everyone. A combined rate, all conversions over all sessions, is a different number that answers how the whole book converts.
Can I benchmark landing-page rates across clients?
Yes, once you match page types. Have the agent pull each client's top pages with get_ga4_landing_pages and tag every page by its job: quote or booking page, service page, or research page such as a cost guide. Then take medians within one tag. On one site, a quote page can convert more than twenty times better than a cost guide, as Hatchwell's do above, so mixing them only measures traffic mix. Use a per-page floor too; 300 sessions in the window is a reasonable start. For what the rate itself counts, see the GA4 integration page.
How often should I rebuild the benchmark?
Quarterly suits most agencies, on a 90-day window. Rerun it on 6 months as a stability check: clients that change quarter between the two windows are noise. Rebuild early when a group gains several clients.
Can I benchmark Google Ads clients the same way?
Yes, at account level. For sites with a Google Ads account linked, DuckDive reads spend, clicks, conversions and cost per conversion. The agent can take medians and quartiles of cost per conversion across clients. It can't split by campaign, because the tool returns an account summary.
Sources
- Mark events as key events. Google Analytics Help. Accessed November 16, 2026.
- What Is the Average Conversion Rate by Ad Platform for Marketing Agencies. AgencyAnalytics blog (February 12, 2025). Accessed November 16, 2026.
- Measures of Location. NIST/SEMATECH e-Handbook of Statistical Methods. Accessed November 16, 2026.
- Interquartile Range. NIST Dataplot Reference Manual. Accessed November 16, 2026.
- Percentiles. NIST/SEMATECH e-Handbook of Statistical Methods. Accessed November 16, 2026.
- What are outliers in the data? NIST/SEMATECH e-Handbook of Statistical Methods. Accessed November 16, 2026.
Last reviewed . Numbers in EXAMPLE DATA blocks are fictional.
