How to Get Your Software Recommended by Gemini

Gemini answers a software question with Google Search behind it. When grounding is on, the model writes its own searches and Google runs them. The answer comes back with the pages it stood on, down to which sentence rests on which page.

That trail is the most detailed any engine in the AI Shortlist Index protocol provides. It is also wrapped in redirect URLs that hide the real page unless you resolve them. This page covers how grounding picks sources for a property-ops question, what the Index has on Gemini so far, and what a vendor should do, in order.

How does Gemini pick its sources?

Google’s Gemini API documentation for grounding with Google Search describes the mechanism. With the google_search tool enabled, the model decides whether a prompt needs a search, and if it does, the response carries a groundingMetadata block.

The block has three parts: the webSearchQueries field holds the searches the model wrote, and groundingChunks the pages they returned. The groundingSupports field maps segments of the answer to the chunks they rely on. Nothing else in the protocol gives you the query the model wrote from a buyer’s prompt.

The redirect problem

The same documentation says the URI on each chunk is a Google-hosted redirect rather than the publisher’s address. The Index’s Gemini adapter is written to resolve every redirect URI to its final URL before classifying the page. The source map will show propertymeld.com, not a Google redirect.

Grounding supports are stored too, so a vendor will see which sentence rested on which page.

Google-Extended

Google’s crawler documentation says the Google-Extended token controls whether a site’s content is used to train Gemini models and for grounding in Gemini. It does not affect Search ranking or AI Overviews. A vendor that blocked Google-Extended in 2024 to keep content out of training may have cut itself out of grounding, so check that first.

What does the Index have on Gemini?

The September 2026 edition’s numbers come from ChatGPT; Gemini carries no mention rate for any vendor until its first edition runs. What the Index does have is ChatGPT: 25 app captures per category, and one retired API capture of the 300-door maintenance prompt whose consulted pages are visible.

That API run consulted 124 pages, 38 of them on reddit.com, 16 on appfolio.com and 12 on buildium.com. It quoted six: two Property Meld pages, AppFolio’s and Buildium’s maintenance feature pages, latchel.com and rentvine.com.

Because grounding runs Google Search, our hypothesis is that Gemini’s chunk set for the same prompt looks more like a Google top ten than like the Reddit-heavy set ChatGPT consulted. PMS vendor pages, listicles, and G2 or Capterra category pages would lead; the first edition with Gemini runs tests that.

A grounding chunk is one page, and a support links a sentence to one or more chunks. The unit of evidence is a page that, on its own, backs a sentence about your product.

What should a vendor fix, and in what order?

1. Fix Google eligibility. Grounding is Google Search, so the page has to be indexed, snippet-eligible, self-canonical and readable without JavaScript. Then confirm Google-Extended is allowed in robots.txt; the AI crawlers guide has the file.
2. Write for the rewritten query. Gemini turns a 300-door maintenance question into searches that read like search queries: maintenance coordination software for property managers, pricing per unit, integrations with Buildium. Title and head your pages the way those searches read.
3. Make each page self-contained enough to be a chunk. A pricing page that says, in one paragraph with the product name in it, “$1.60 per unit per month, $160 monthly minimum” supports a sentence by itself. A page where the product name is in a header image and the price loads by script does not.

Then corroboration and consistency

4. Get corroborated. A sentence about your product that rests on two chunks from two domains is one we expect the model to keep. Think of your pricing page plus a Reddit thread or a marketplace listing. A sentence resting only on your own site is one we expect it to hedge.
5. Make the entity agree with itself. One name, one category and one integration list across your site, Crunchbase, LinkedIn, G2, Capterra and the PMS marketplaces. Grounding pulls chunks from several of those at once, and the entity consistency guide covers the audit.
6. If you monitor Gemini yourself, resolve the redirects. Without that step your source report is a column of Google URLs, and you will not know whether the page behind the mention was yours.

The category source map lists the corroborating domains with counts and will add Gemini’s resolved domains once runs exist. The prompt library has three integration prompts per category to write those pages against.

What does the Audit find?

The Audit is $3,500 and takes seven days. It runs the full 25-prompt set for your category in ChatGPT, three runs each.

It returns your position, plainly, and the cause, as ranked findings labeled verified, inferred or hypothesis with evidence links. Then the source gap, a 90-day fix scoped to a Sprint, and what we won’t promise: rankings, traffic, leads or revenue.

Most of the fixes above carry across engines.

When the Index runs Gemini, the Audit will add the searches the model wrote for each prompt and the resolved pages it grounded on. It will also show which sentences about you rested on which page. If you want to talk first, book a 20-minute call.

The mechanism in more depth is in the Gemini grounding guide. The other engines have their own pages: getting your software recommended by ChatGPT, earning Perplexity’s citations, and appearing in Google AI Overviews.