How fast does GetIntent swap landing-page copy? A local benchmark
On a local test page, the rewritten headline and CTAs were on screen a median of 69 ms after navigation on the first visit for a new keyword, not counting the AI model's own generation time, 58 ms when the copy was already cached on the server, and 5 ms on a reload in the same tab. Visits without campaign parameters were never hidden and never rewritten. On WordPress with the snippet in WPCode the same steps took 145, 121 and 47 ms. 30 runs per scenario on November 22, 2026; method, script and raw data below.
- Published
- November 22, 2026
- Author
- Andreas Hatlem, founder of GetIntent (Getia AS)
- Sample
- 30 runs per scenario, 4 scenarios, plain HTML and WordPress
- Environment
- Local production build, November 22, 2026
Key findings
- 1
Visits without campaign parameters were left alone: hidden for 0 ms at most, no rewrite request (0 at most) and no model call. Organic and direct visitors see your page as it is.
- 2
First visit for a keyword with no cached copy: the rewritten copy was on screen a median 69 ms after navigation (90th percentile 702 ms). This is GetIntent's own work (script, config, cache lookup, the rewrite request and the swap); the AI model's generation time comes on top and was not part of this measurement.
- 3
When the copy for that keyword was already cached on the server, a new visitor in a new tab saw it after a median 58 ms (90th percentile 404 ms), with 1 model calls across all those runs.
- 4
A reload in the same tab is served from sessionStorage: a median 5 ms, with no rewrite request.
- 5
On a first visit the install snippet's guard hid the page for a median 63 ms before the swap. It never hides a page for longer than 800 ms: a rewrite that is not ready by then appears after the original wording.
- 6
WordPress WordPress 7.1.5, snippet in WPCode 2.4.2 "Header" field, PHP built-in server: the same scenarios took 145 ms (new keyword, 90th percentile 554 ms), 121 ms (cached) and 47 ms (reload). The difference to plain HTML is mostly WordPress itself: its page reached DOMContentLoaded after a median 59 ms, against 6 ms for the static page.
Results by platform and scenario
Time from navigation start to the rewritten copy being on screen, and how long the guard kept the page hidden.
| Platform | Scenario | Rewritten | Median | 90th pct. | Hidden (median) | Model calls |
|---|---|---|---|---|---|---|
| Plain HTML | No campaign parameters | 0/30 | – | – | 0 ms | 0 |
| Plain HTML | New keyword (not cached) | 29/30 | 69 ms | 702 ms | 63 ms | 29 |
| Plain HTML | Cached on the server, new tab | 30/30 | 58 ms | 404 ms | 53 ms | 1 |
| Plain HTML | Reload in the same tab | 29/30 | 5 ms | 6 ms | 2 ms | 0 |
| WordPress | No campaign parameters | 0/30 | – | – | 0 ms | 0 |
| WordPress | New keyword (not cached) | 30/30 | 145 ms | 554 ms | 100 ms | 30 |
| WordPress | Cached on the server, new tab | 30/30 | 121 ms | 253 ms | 79 ms | 0 |
| WordPress | Reload in the same tab | 30/30 | 47 ms | 65 ms | 5 ms | 0 |
Model calls come from a local stand-in for the AI gateway that answers at once, so 'new keyword' times exclude model generation. Localhost has no network latency: on a real site, add your visitors' round trips to getintent.co (two requests on a cache miss, one on a server cache hit, none on a reload in the same tab).
What setup takes
Going live on an existing page is six steps. We timed the product side of them, not a person doing them:
- Create a free account and add your site's domain.
- Copy the install snippet from Dashboard > Install.
- Paste it into the page head: in WordPress with the WPCode plugin (Code Snippets > Header & Footer > Header), in Webflow under Site settings > Custom code > Head code, in HubSpot in the site header HTML setting for your pages.
- Publish the page.
- Optionally choose which elements may change. By default the script rewrites up to 50 text elements per page view; text in the header, navigation and footer is left alone.
- Add the Google Ads final URL suffix from Dashboard > Install, or tag a link with the Intent Builder, and open it.
Measured: on a local WordPress WordPress 7.1.5, snippet in WPCode 2.4.2 "Header" field, PHP built-in server, after the snippet was saved in the WPCode Header field, the first tagged visit showed rewritten copy a median 145 ms after navigation, model time excluded. Nothing else had to be configured for that first rewrite.
Not measured: Webflow and HubSpot host your pages, so they cannot be run locally; the steps above come from our install docs, and the snippet is the same. We have not timed people doing the steps, so we publish no setup duration in minutes.
What this means for your pages
GetIntent's own share of the swap is small: tens of milliseconds on this machine. On a real site the visible delay on a first visit for a new keyword is mostly the AI model's generation time plus network round trips, which is why the guard caps hiding at 800 ms and falls back to your original wording.
Repeat traffic is the fast path: once a keyword's copy is cached, every later visitor gets it without a model call, and a reload in the same tab needs no request at all.
Methodology
Environment: a production build of GetIntent (next start) on localhost with its own PostgreSQL and Redis, Darwin 24.6.0 arm64, 10 CPU cores, Node v22.23.4, Chromium 148.0.7778.96 (Playwright, headless). The AI gateway was replaced by a local stand-in that returns rewritten copy at once, so model latency and model output quality are not part of the result.
Test pages: a static landing page with a header, a hero (headline, subheading, two buttons), three feature blocks, a closing section and a footer, served by a local HTTP server; and the same content as a WordPress page with the snippet in the WPCode Header field. The install snippet is the one from Dashboard > Install, guard included. The static page is downloadable below.
Procedure: for each platform, three warm-up visits, then 30 runs of each scenario in a fresh browser context: a visit without parameters, a visit with a Google Ads final URL suffix (gi_*) for a keyword never seen before, a reload of that tab, and a new tab with the same keyword. The timer starts at navigation (performance.now() is 0) and stops when the script fires its getintent:transformed event.
Definitions
- Time to rewritten copy
- Milliseconds from navigation start to the getintent:transformed event, which the script fires right after it writes the new text into the page.
- Hidden
- How long the guard's class was on the html element, which makes the body transparent. Zero when the guard never applied, as on visits without campaign parameters.
- Rewrite requests
- Requests to /api/rewrite and /api/rewrite/check during the visit.
- Model calls
- Calls that reached the AI gateway stand-in, counted on its side.
Limitations
- Localhost: no network latency, so real-world times are higher by your visitors' round trips to getintent.co.
- The AI model's generation time is excluded on purpose. It depends on the model, page length and load, and is the largest part of a first visit for a new keyword.
- One machine, one browser (headless Chromium), one page design. Heavier pages and slower devices take longer to load before the script runs.
- WordPress ran on PHP's built-in server with SQLite, slower than a tuned host; Webflow and HubSpot were not measured.
- In this run a new tab with a cached keyword was served by the regular rewrite request, not the faster cache-check request, because Google Ads keywords are normalized differently on the two paths. A fix would make that scenario faster, not slower.
- The machine was busy with other work during the run (one-minute load average 13.1 at the start and 29.8 at the end, on 10 cores). That explains the long 90th percentiles; read the medians as the typical figures.
Check or reproduce the results
The benchmark script (scripts/research/run-swap-latency-benchmark.ts in the GetIntent repository) seeds a site in a local database, starts the test page and the gateway stand-in, drives Chromium with Playwright and writes the summary and CSV below. It refuses to run unless the database and the app are on localhost. Ask for it at [email protected].
Download the data
- results.csv: one row per visit with platform, scenario, keyword, timings, requests and model calls
- summary.json: environment, run timestamps and every aggregate figure
- test-page.html: the static test page with the install snippet
Free to cite with a link to this page.
How to cite this study
GetIntent (Getia AS), “How fast does GetIntent swap landing-page copy? A local benchmark”, November 22, 2026, https://getintent.co/research/swap-latency-2026
Last updated November 22, 2026. Data collected November 22, 2026.