Grounding

What does a grounded answer look like?

It shows its reading. On 3 October 2026 ChatGPT answered "how does AI pick which sources to cite", a question we track for attensira.com, and we read the stored answer over the Attensira MCP. The table at the top of this page lists what it grounded on: five domains in its sources consulted, all of them OpenAI's, and two of those cited inline next to the sentences they support.

The answer explained the gap itself: "OpenAI explicitly notes that the Sources list can contain more URLs than the inline citations because only the most relevant references are normally surfaced as citations." Grounding has two outcomes for a page, then. It can be read, and it can be credited, and the second is the smaller set.

For a publisher that difference matters more than the definition. A page in the consulted list shaped the answer without anyone seeing its name. A page cited inline is attached to a claim a reader can click. When an engine answers without grounding at all, there is nothing to credit; AI hallucination shows what that looks like on our own brand.

What is grounding?

Supplying a language model with source material at answer time and having it build the answer from that material, with citations. The sources are usually:

• web search results, as in ChatGPT search, Perplexity, Claude with web search, Gemini and Google's AI Overviews;
• private documents, as in enterprise assistants that search a company's files.

Google's Gemini API documentation describes grounding with Google Search as connecting the model to real-time web content and letting it cite verifiable sources beyond its knowledge cutoff. The technique underneath is retrieval-augmented generation.

How does grounding work?

1. The model decides to search. For many questions it does; for some it answers from memory and attaches no sources.
2. It writes queries and retrieves results. The search step returns pages; passages are selected and placed in the model's context.
3. It writes from what it read. The answer draws on the passages and attaches some of them as links or footnotes: in our ChatGPT example, two of the five domains it consulted. See source citation.

Each step is a place your page can drop out: not found, not selected, or used without being named.

What does grounding fix, and what not?

It fixes staleness. A grounded answer can carry facts newer than the model's knowledge cutoff, such as a new price, a rebrand or a launch.

It reduces, but does not remove, hallucination. The model has real text to work from, so it invents less. It can still misread a passage, draw from one that answers a neighbouring question, or blend two sources into a sentence neither said. A link shows what was read, not that the sentence is right.

And it introduces a new dependency: the answer is now only as good as what the search step found. If your page was not fetchable, an outdated third-party page may be grounded on instead.

What are the objections about grounding?

"Grounded answers are always accurate." Check the sentence against the page. Grounding moves the risk from memory to reading.

"Our brand is in the training data, so grounding doesn't matter." Training is frozen at the cutoff. Anything you changed since only arrives through grounding.

"We can't influence what the model searches." You can influence what it finds: be fetchable by each engine's search agent, and have a page that answers the question in the words people use.

"Grounding means the model trusts our site." It means your page was retrieved for this question. Trust is earned passage by passage.

How does Attensira work with grounding?

Attensira stores every answer each engine gives to your buyers' questions, with the links it attached and the sentence each link supports, so you can see when answers were grounded in your pages, in someone else's, or in nothing. The crew then makes your pages fetchable and quotable and gets you onto the sources engines ground in, shipping each change after your approval. The table at the top shows the gap that work closes: five domains read, two credited. Citation tracking covers how each cited link is recorded, and off-site work how Attensira earns a place on the sources engines ground in.