The websites linked in 1,000 AI software-buying answers

QuickBooks, Salesforce, Sage and Shopify occupied the first four places in our hostname count. Links to quickbooks.intuit.com appeared in 109 answers; salesforce.com in 90; sage.com in 82; and shopify.com in 81.

We asked 100 software-buying questions ten times each through the Gemini 2.5 Flash-Lite API, using Parallel Turbo web search. The questions covered ten categories and explicitly asked for source links where available. We then counted the URLs in the answers.

Across the 1,000 usable answers, we found 949 distinct normalized URLs on 521 hostnames. 994 answers contained at least one link. These figures describe this API experiment, not the usual citation rate of a consumer AI assistant.

On September 28, we extended the source review to the 50 most frequent hostnames. Forty-three belonged to software or service vendors, four to technology publishers, and one to an app marketplace. Two remained unclassified because their primary pages were unavailable. A vendor link does not necessarily recommend that vendor.

These 50 hostnames appeared in 828 answers and accounted for 1,568 of 3,966 answer–hostname appearances. Repeated links to the same hostname within one answer count once. The 800 figure is therefore a lower bound for answers containing a vendor site in this corpus. The other 471 hostnames were outside this supplement.

Vendor-written comparisons and buying guides were part of the information linked alongside these answers. Clear prices, intended customers, conditions and comparison criteria help a buyer inspect those pages. This study did not test whether editing them increases AI visibility. Classification used AI-assisted primary-site review, without independent human double-coding.

In answer crm-02-01, we asked for CRM options for a 50-person B2B sales team with a long sales cycle and detailed forecasting needs. The response linked to Salesforce’s B2B sales-tools page beneath both Zoho CRM and Microsoft Dynamics 365 Sales. It also linked to a HubSpot forecasting article beneath Zoho CRM and BoostUp.

The name of the website and the name of the recommended product were different. A source report that counts only a brand’s own domain would miss that relationship. A report that treats every vendor link as a recommendation of that vendor would misread it.

This example shows where links were placed. We did not verify every product claim against the linked pages, and the link does not prove that the page caused the recommendation.

A payroll buyer and an analytics buyer do not produce the same set of links. We gave every category the same number of answers so the category comparison would be readable. That balance is a research choice; it does not reflect market demand.

We counted hostnames separately. For example, hubspot.com and blog.hubspot.com are two hostnames in this dataset, even though they belong to the same brand. The chart measures the breadth of the visible link set, not the number of independent publishers.

For a marketer, the useful next step is to narrow the view to the buying situations that matter to the business. A site that appears frequently in the pooled list may have little presence in your category. The evidence page keeps the original questions and their ten answers together.

The provider returned 5,000 normalized source-URL appearances across the main sample, counting a URL once per answer. Of those, 4,399 also appeared in the corresponding answer text: 88.0%. The same page appearing in ten answers contributes ten appearances to these totals.

The remaining source URLs were present in the provider metadata but absent from the answer’s text. We kept those URLs in a separate field. Counting all of them as visible citations would overstate what a reader could see in the answer.

An answer link can also be a product destination rather than a citation supporting a claim. We use “answer links” for that reason. This audit follows visible URLs; it does not trace training data, hidden reasoning, or causal influence.

We chose Parallel Turbo, requested up to five results, and used a plugin whose default instructions tell the model to incorporate and cite search results. Different search settings may produce different links.

I would show the recommended company and the linked website in separate columns. The Salesforce example explains why. Next to them, I would keep the buying question, the answer date, and the exact page URL so a reviewer can inspect the context.

For content planning, I would read the recurring pages before deciding what to publish. A comparison guide, a product page and a support article answer different needs. A hostname count alone cannot tell you which of those formats is appearing or whether its claims are accurate.

If we publish a comparison, I want it to help the buyer make the decision in the question: explain the relevant tradeoffs, show where the information comes from, and make the scope clear. This study does not prove that copying a frequently linked format will earn citations.

I would then monitor the same relevant questions over time and inspect changes in the answers. One appearance is a useful lead for investigation; it is not evidence that an SEO change worked.

We wrote ten brand-free questions for each of ten software categories, then requested ten fresh answers per question. Each request asked for five options, brief explanations and source links where available. All questions were in English. The Japanese edition reports the same corpus rather than a separate Japanese experiment.

The requests used google/gemini-2.5-flash-lite through OpenRouter, temperature 0.2, a 1,200-token output limit, and reasoning disabled. There was no system prompt or conversation history supplied by us. The web plugin used Parallel, turbo mode, and up to five results. Five sequential batches ran concurrently, with the category order fixed inside each batch.

The original collection took place on September 23, 2026. Four responses returned an HTTP success status but an error finish reason and incomplete text. We excluded them and collected four replacement answers separately. The released corpus contains 1,000 usable answers from 1,004 provider responses, with the failures retained in the research records. Three earlier pilot answers are excluded.

URL extraction is deterministic, with AI-assisted checks. We normalize hostname casing, remove leading www, fragments, trailing slashes and specified tracking parameters. Distinct paths remain distinct. Source-type notes come from a limited review of selected sites; the headline counts do not depend on assigning every site a type.

The repeats share questions and may share cached search results. They are not 1,000 independent buyer situations. We did not test other models, consumer apps, languages, personalization, purchase outcomes, or the effect of a website change.

Provider-reported charges were $1.4549 for the main sample, including the four failed responses and their replacements, plus $0.0043 for the excluded pilot. Five earlier network attempts returned no response or billing metadata; we retained a separate $0.75 budget reserve rather than treating their cost as zero.

The evidence page groups all 1,000 usable answers by question. It shows the answer text, the links we extracted, and the provider-reported source URLs separately. You can download the data and check the counts behind the charts.