CustomGPT.ai Review 2026

CustomGPT.ai Review: We Indexed 2,534 Of Our Own Pages Into It

Most reviews of a retrieval-augmented chatbot builder describe the feature list. This one describes what happened when we put a real corpus into it and asked questions we already knew the answers to — which is the only way to find out whether a tool that cites sources is citing them correctly.

NorwegianSpark runs two live agents on CustomGPT. The first indexes 2,534 pages and 4,955,173 words across our own websites, and exists to answer one question: which of our pages already cover a topic, so we do not write the same thing twice. The second indexes 252 pages of Google’s Search Central documentation, so that questions about Google’s published policy get answered from Google’s own wording rather than from folklore. Both were built and tested in the account, not in a sandbox.

What it is

CustomGPT builds a chat agent on documents you supply, and answers only from those documents. Sources can be a sitemap, a list of URLs, uploaded files (PDF, ZIP, audio), Google Drive, SharePoint, OneDrive, Notion, Confluence, Zendesk, YouTube and Vimeo, an RSS feed, or a WordPress, Shopify, Wix or Ghost site. Deployment is an embed, a live-chat widget, WhatsApp, Slack, Intercom, or an MCP server — that last one meaning the agent can be queried from a coding assistant rather than through the web UI.

Ingestion: fast, and it took a large corpus without complaint

The 252-page Google documentation set went from an empty agent to fully indexed in a few minutes. The 2,534-page fleet index is comfortably inside the Standard plan’s 5,000-documents-per-agent ceiling.

One practical wrinkle worth knowing before you start: the “URL List” button in Add Source is not a source type. It opens a separate helper that converts a pasted list into a hosted sitemap, which you then add as a Sitemap source. It works, and it accepted 252 URLs in one paste, but the extra hop is undocumented in the interface itself and the resulting sitemap is hosted on a third party’s storage rather than yours.

Retrieval: accurate on what it returns, incomplete on what it finds

Asked which of our pages target a specific keyword, the fleet agent returned five URLs and named the site that owns the query. We checked every one against the live sitemap: all five were real pages that returned HTTP 200. Asked in an earlier test to list pages about a subject we publish nothing on, it answered “I don’t know the answer” and invented nothing — the failure mode that matters most in a tool like this.

The limit is recall, not precision. In the same test it missed two further pages that also compete for that keyword. Treat a short answer as a subset, never as proof that nothing else exists. That distinction is the difference between using this tool well and drawing a false conclusion from it.

The flaw we found: a correct quotation attached to the wrong page

This is the finding we would want to read before buying, so it goes in the review rather than in a footnote.

We asked the Google-documentation agent what Google requires of an affiliate page for it to count as adding value. The answer was substantively right, quoted Google’s own wording, and correctly distinguished a requirement from a recommendation. It then cited a specific documentation URL as the source of that sentence. We fetched the page and searched it: the quoted sentence is not on it.

So the citation is a lead, not a receipt. If you are using this to settle a question of fact — a compliance question, a policy question, anything you would act on — open the page it names and confirm the wording is actually there. The substance was sound; the pointer was not.

The trap nobody documents: the persona freezes at whatever you index first

CustomGPT auto-writes an agent persona from the content present when the agent is created. Adding more sources later does not update it.

Our fleet agent was created when only one website had been indexed, and it wrote itself a persona describing that one site. Twenty-one sites later it was still refusing anything outside that first subject — a question about VPN pages came back as “that falls outside my scope.” The documents were there the entire time. The symptom looks exactly like broken retrieval, and no amount of re-indexing fixes it, because the problem is a text box on the Persona tab.

The auto-generated starter questions have the same defect and are easy to miss, because they sit on a different tab. Both ours still described the first site indexed, days after the corpus had grown past it.

If your agent starts refusing a whole subject area, read the persona before you touch the sources.

The interface is the weakest part

Over one working session the app became unresponsive several times, one settings panel never finished loading at all, and a save button reported nothing while silently failing — the setting simply was not there on reload. Nothing was lost, but you cannot trust a save here without leaving the page and reading the value back. That is a real operational cost on a product at this price.

Pricing

Standard is $99/month, or $89/month billed annually: 10 agents, 1,000 queries per month, 5,000 documents per agent, 3 team members. Premium is $499/month ($449 annual) for 25 agents, 5,000 queries, 20,000 documents per agent and 5 seats. Enterprise is quoted on request. Prices read from CustomGPT’s published pricing page at the time of writing; check it before purchase.

Two things the plan boundary decides. Automatic re-syncing of online sources is a Premium feature — on Standard, a site that changes has to be re-synced by hand, which matters a great deal if the corpus is your own documentation. And a query is not a message: a single question consumed several credits in our account, so the 1,000-per-month figure is not 1,000 questions.

Who it is for

It earns its price where you have a corpus you already own, and the cost of a wrong answer is high enough that “answer only from these documents, or say you do not know” is worth paying for. Internal documentation, a support knowledge base, a large publishing archive you need to search by meaning rather than by keyword.

It is a poor fit if your corpus is small enough to paste into a general assistant, if you need the answer set to be exhaustive rather than indicative, or if you need every citation to be exact without checking. On our evidence it is a strong research aid and a weak system of record.