Fannie Mae and FHA publish their rules in fundamentally different ways.
The Selling Guide is a website. Every topic has its own page and its own address, so B3-4.3-04 on personal gifts is a thing you can link to, quote from, and come back to. Handbook 4000.1 is a single PDF of roughly 1,900 pages, and a section inside it has no address of its own — it has a position in an outline, II.A.1.b.iii, which is not something you can send anybody.
I wanted to know whether that difference changes the answer you get from an AI assistant. So I tested it.
The test
Eight underwriting concepts, each asked twice: once about FHA, once about Fannie Mae. Same concept, same phrasing, only the rulebook different.
gift funds · self-employment income · bankruptcy waiting period · condo project review · property condition · rental income · monthly debts and DTI · occupancy
Each pair went to three assistants — ChatGPT, Claude and Perplexity — and for every answer I recorded three things: whether the agency’s own current site was cited, whether an archived or superseded copy was cited, and how many distinct sources the answer drew on.
Forty-eight answers in total. Eight per cell is a small sample and I will not pretend otherwise, so what matters below is where all three engines moved the same direction, not the size of any single number.
What came back
| Engine | Rulebook | Cited the agency’s own site | Cited an archive | Sources per answer |
|---|---|---|---|---|
| ChatGPT | FHA | 8 of 8 | 2 of 8 | 3.2 |
| ChatGPT | Fannie | 8 of 8 | 0 of 8 | 2.4 |
| Perplexity | FHA | 7 of 8 | 0 of 8 | 12.2 |
| Perplexity | Fannie | 8 of 8 | 0 of 8 | 8.5 |
| Claude | FHA | 3 of 8 | 0 of 8 | 9.6 |
| Claude | Fannie | 7 of 8 | 0 of 8 | 7.8 |
Three things fall out of that.
Every engine used more sources to answer the FHA question. Perplexity 12.2 against 8.5, Claude 9.6 against 7.8, ChatGPT 3.2 against 2.4. Different engines, different architectures, same direction every time. When the rule is inside a PDF, the assistant goes looking in more places.
ChatGPT cited HUD’s archive twice, on gift funds and on condo project review. Both times alongside current hud.gov, so the answer blended live and archived guidance without distinguishing them. On Fannie it never happened, in any engine.
And Claude did not reach HUD at all on five of eight FHA questions. That is the result I did not expect.
Where the FHA answers actually came from
When Claude could not get to HUD, it answered from these instead:
fha.com · fhahandbook.com · fhaloans.com · fhanewsblog.com · gustancho.com · lendingtree.com · creditkarma.com · chase.com · griffinfunding.com · neighborsbank.com · bayou-mortgage.com
Those are lender marketing sites and comparison shopping pages. Some are perfectly accurate. None of them is the Handbook, none carries an effective date you can rely on, and none tells you when it was last checked against the current text.
On the same eight concepts asked about Fannie, the same assistant reached Fannie’s own site seven times out of eight.
The contrast is not about which agency writes better rules. It is about which agency’s rules an assistant can actually get to.
Why the format does it
A retrieval system wants a page: a unit of text with an address, about one topic, that can be fetched and quoted.
Fannie gives it exactly that. One topic, one page, one URL, and the section number is in the address. An assistant asked about gift funds finds the gift funds page and cites the gift funds page.
FHA gives it a 1,900-page PDF. To answer a question about gift funds, an assistant has to either ingest the whole document and locate the passage, or find something that has already done that work. The second is easier, and the things that have already done it are lender blogs and, sometimes, an older copy of the handbook that happens to be more reachable than the current one.
There is a fairness point to make here. HUD’s content sits across several subdomains — hud.gov, archives.hud.gov, glb.hud.gov, answers.hud.gov, entp.hud.gov — and some of those are systems rather than document mirrors, so this is not simply a badly built website. The handbook is also a genuinely hard thing to publish: it is enormous, heavily cross-referenced, and amended by Mortgagee Letters between reissues. But the consequence stands regardless of the cause. A rulebook that cannot be addressed section by section gets summarised by somebody else, and the summary is what the assistant reads.
What this means if you are relying on the answer
Three practical consequences, in order of how much they should bother you.
An FHA answer is more likely to be assembled from secondary sources. Not wrong, necessarily. But an answer built from five lender blogs is not the same artifact as an answer built from the Handbook, and nothing in the response tells you which one you are holding.
Archived guidance reads exactly like current guidance. archives.hud.gov is a hud.gov address. It looks official because it is official. It is simply not in force. An assistant citing it alongside the live site produces something that appears well sourced and may describe a requirement that changed years ago.
The failure is silent and it is topic-dependent. The same assistant that answered the Fannie question from Fannie’s own page answered the FHA question from a mortgage broker’s blog, in the same session, with the same confidence and no change in tone. You cannot tell from the answer which one you got.
What to do about it
Ask for the source, every time, and look at it. Not “cite your sources” as a formality, but actually reading the host. If an FHA answer is resting on fha.com and lendingtree.com, you have learned something important about how much weight it will carry.
Watch specifically for the archive. For HUD, archives.hud.gov is the tell. If it appears, the answer may be describing superseded guidance, and it will not say so.
Give the assistant the text instead of making it go find the text. This is the whole point. An assistant that already has the current section in front of it does not need to choose between HUD’s PDF, HUD’s archive and somebody’s blog. The retrieval problem disappears, because there is no retrieval.
That last one is why I publish this library. FHA Handbook 4000.1, the Fannie and Freddie guides, VA, USDA, Ginnie Mae and the CFPB regulations — 3,127 regulations, 7,685 obligations, each one verbatim, section by section, with its own address and the date it was captured.
Download the corpus from Claude for Compliance and run it locally: Excel · CSV · source files · or browse it online first.
Free, ungated, nothing to install and no licence to accept. Load it into a Project alongside your own policies and every answer is grounded in a copy you control, on your side of the firewall, that does not change under you between sessions.
The honest caveats
Eight questions per cell is a small sample. What I trust here is the direction, which was consistent across three independent engines, not the precise percentages. I intend to re-run it on a wider set.
This is a snapshot. Retrieval behaviour changes as these systems change. A result from August 2026 is a result from August 2026.
No claim that FHA’s requirements are worse, or that Fannie’s are better. This is about publication format and what it does to machine retrieval, nothing else.
My library has its own limit, and it is the same class of problem. It mirrors published text and stamps every page with the date captured. It does not yet tell you whether a Mortgagee Letter has amended that section since. Confirm at the source before you rely on it. That is the honest state of it, and it is the next thing I am building.
Everything here is a working draft for your own compliance and legal review, not legal advice.