Reaching Visitors Who Do Not Read English: An AI Translation Implementation for Visit Santa Cruz County
A destination markets itself to people who arrive from everywhere, and a share of them do not read English. We implemented Pare Linguist — the AI translation product we operate — on the site Visit Santa Cruz County already had.
The work
The Challenge: An Audience That Arrives From Everywhere
Visit Santa Cruz County markets a California coastal destination to people who come from everywhere, and a share of them do not read English. A visitor who cannot read the site cannot use it to plan a trip — and the three available ways of fixing that, human translation, static PDFs and a browser widget, each fail either at scale or at the point where the words have to be right.
The organization needed translation on the site it already ran, not a second multilingual site to build and keep current beside it.
The Strategy: Deploy the Product on the Site the Organization Already Had
Implemented Pare Linguist over the existing site — a deployment of infrastructure we operate, not a translation project the client funds and maintains.
- A subscription, not a build. Real translation infrastructure — engine integration, editorial workflow, QA, governance tooling — priced as a deployment of a multi-tenant product rather than as bespoke work one organization pays to build alone.
- Translation happens close to the reader. A visitor who selects Spanish gets the translated page without a round trip to the origin, so translation never becomes a performance liability on the organization’s own infrastructure.
- The organization holds the dictionary. The same English string always produces the same output, nothing is translated twice, and staff can audit any entry, set preferred terminology, protect the strings that must never be translated and correct the one word that matters — which on a destination site is usually the name of a place or a business.
The Outcomes: Live on the Site a Destination Already Ran
In production at Visit Santa Cruz County, on the site the organization already had.
- Intelligence | Live AI translation at Visit Santa Cruz County | A reviewable dictionary as the record, with the model populating it
- Trust | Staff can audit and override any translation | Governance that keeps the names of places and businesses exact
The detail
Why the existing options fail
- Human translation does not scale. High quality, but slow and expensive, and fundamentally incompatible with the content velocity modern institutions produce. By the time an emergency notice is translated, the English source has often updated again.
- PDFs make the translated audience second-class. Static, hard to update, often behind a link no one finds. The non-English reader encounters the institution as a slower, less complete version of the experience everyone else gets.
- Browser widgets fail exactly where it matters. Inconsistent quality, no preservation of formatting or technical terms, no governance over institution-specific language and no institutional ownership of the result. Language access becomes the reader’s problem to solve rather than the institution’s to provide.
- Building it bespoke costs too much to justify. Real translation infrastructure needs engine integration, editorial workflow, QA, governance tooling and delivery infrastructure. Built once for one institution, the price is high enough that most cannot justify it for what their stakeholders file under overhead.
How the implementation works
- A product deployed, not a project funded. Most agencies would build translation integration as bespoke client work and hand off operations. Pare Linguist is multi-tenant infrastructure we operate, so the organization pays for a deployment while the cost of building the engine integration, workflow, QA and governance tooling amortizes across every institution running it.
- Translation happens close to the reader. A visitor who lands on a page and selects Spanish gets the translated page without a round trip back to the origin. Latency stays low, and translation never becomes a performance liability on the organization’s own infrastructure.
- A stored dictionary, not a call per request. Translations are kept after first generation, which buys three things at once: nothing is translated twice, the same English string always produces the same output rather than drifting per request, and the dictionary is reviewable and editable by the organization’s own staff. The dictionary is the source of truth; the model populates it.
- Human override where one word has to be exact. Staff set preferred terminology once, protect the strings that must never be translated — business names, place names, addresses — and correct anything the model gets wrong. On a destination site those strings are most of what a visitor is looking for, so fixing one specific word matters more than aggregate quality.
- Drop-in, not a parallel CMS. It runs on the site the organization already had, rather than requiring a second multilingual site to stand up and maintain.
What is and is not claimed
- No measured outcomes, and that is deliberate. The client does not currently surface audience reach, content volume or accuracy metrics. This page writes to structural evidence — the deployment exists and the architecture is operating — rather than to numbers no one has published.
- One named deployment is the honest framing. As more come online the pattern claim gets stronger. For now it is Visit Santa Cruz County, named.
AI translation for institutions questions
Why not just use a browser translation widget?
Because it fails exactly where an organization is accountable for what its own site says: inconsistent quality, no preservation of formatting or technical terms, no governance over institution-specific language and no institutional ownership of the result. Language access becomes the reader’s problem to solve rather than the institution’s to provide.
How does AI website translation stay accurate on content that matters?
With a stored dictionary and human override. Translations are kept after first generation, so the same English string always produces the same output — and institutional staff can audit the dictionary, set preferred terminology, protect strings that must never be translated and correct anything the model gets wrong. When a place name or a business name has to be right, fixing one specific word matters more than aggregate quality.
Why do institutions not build translation infrastructure themselves?
The price. Real translation infrastructure needs engine integration, editorial workflow, QA, governance tooling and delivery infrastructure — built once for one institution, that costs more than most can justify for what their stakeholders file under overhead. Pare Linguist is multi-tenant infrastructure we operate, so the build cost amortizes across deployments instead of landing on each client.
Does AI translation slow the website down?
Not in Pare Linguist’s architecture. Translation happens close to the reader — a visitor who selects Spanish gets the translated page without a round trip to the organization’s origin — so latency stays low and language access never becomes a performance liability on the institution’s own infrastructure.
Where is Pare Linguist in production?
At Visit Santa Cruz County, the destination marketing organization for the California coastal county. The deployment runs on the site the organization already had; it required no parallel multilingual CMS.
Client leadership
Christopher Murray
Founder & CEOChris Murray founded Pare & Co in 2007 and rebuilt the firm in 2026. He led the work behind Pare Linguist, the translation product this implementation runs on.