Earnings season without drowning: comparing four quarters of the same call
Listening to one earnings call tells you what a company said. Reading four of them side by side tells you what changed. The second thing is much harder to do by ear, and much easier to do once the audio is text.
What this article is about
This is an information-handling workflow — how to capture, search and compare a body of recorded material without losing your afternoon to it. It is not investment advice, it contains no view on any company or security, and nothing here is a claim that better-organised information produces better returns. We build transcription tools; research method is the part of this we can speak to honestly, so it is the only part we do.
The mistake is treating a call as an event
The default way to cover earnings is to attend. You get the calendar, you block the hour, you listen live, you take notes, and then the next one starts. Twelve companies into a season the notes are inconsistent, the early ones have gone cold, and you could not reliably answer a question as simple as did they use the word "transitory" last quarter too?
The problem is not effort. It is that listening is a linear pass over a document you can only read once, at the speed the speaker chooses, in the order they chose. Nothing you do during that hour makes the hour repeatable.
The alternative is to stop treating a call as an event to attend and start treating it as a document to add to a set. A set can be searched in one query. A set can be compared. An event can only be attended.
The workflow
This runs on one company across four quarters. It works the same way on one quarter across a peer group; the four-quarter version is the better place to start because a company is a more controlled comparison than a sector — same business, same speakers, mostly the same script.
-
Define the set before you collect anything
Write down the four specific calls you are going to work with, by company and quarter, before you open a single page. This sounds like bureaucracy and is actually the whole discipline: a defined set has an edge, so you can finish. An undefined set — "I'll go through recent calls" — has no edge, so you graze until you are tired and call that research.
-
Collect the filed text and the spoken audio separately
These are two different sources and it matters that you keep them apart. The earnings press release is commonly furnished to the SEC as an exhibit to a Form 8-K under Item 2.02, Results of Operations and Financial Condition. That document is free, permanent, and full-text searchable on EDGAR. The call — prepared remarks plus analyst Q&A — is usually a separate event that most companies do not file. That gap is precisely why the audio is worth the trouble: the numbers are in the release, but nobody is asked a question in a press release.
The call itself lives on the company's investor relations page as a webcast. Availability varies and there is no rule that fixes it: some IR archives go back years, some post a replay for a limited window, and some put it behind a registration form. You cannot tell which kind of page you have until you need the file and it is gone.
-
Capture on the day, not when you get to it
Because replay windows are unpredictable, the habit that saves you is capturing the audio the day it happens rather than trusting it will still be there. For a webcast playing in a browser tab, the Whipscribe Chrome extension records that tab's audio and sends it straight to a transcript — which is the only mechanism that works when the source is a streaming player rather than a file you can save.
Where a direct media URL does exist, paste it and skip the recording step entirely. Where the source is DRM-protected — Spotify is the common case — no tool can fetch it, ours included, and we tell you that before you wait rather than after. Do not go looking for a way around a platform's protections; find the same content where it is published openly instead, which for earnings material it usually is.
-
Transcribe the whole set in one sitting
Four calls is roughly four hours of audio, and the single biggest reason this workflow fails is that people transcribe one, read it, and never do the other three. Queue all four before you read any of them. The comparison is the deliverable; a single transcript is just a call you attended slowly.
Ask for timestamps and, where the tool offers it, speaker labels. On an earnings call the speaker labels are doing real work — a question answered by the CFO and the same question answered by the CEO are different observations, and an unlabelled wall of text loses that permanently.
-
Write your term list before you read a word
Decide what you are looking for before you start reading, and write it down. This is the step people skip, and skipping it is what turns comparison into confirmation: read four transcripts with no list and you will find whatever you already believed, because you will notice the sentences that match it.
A term list is fifteen to thirty strings. Some are specific to the business; some are structural and travel across every company you will ever look at.
-
Run the list across all four, and log every hit with its quarter
One term at a time, all four transcripts, hits recorded with quarter and timestamp. Tedious, mechanical, and about twenty minutes — which is the point. It is the part of research that is legitimately automatable by search, and doing it by memory is what memory is worst at.
-
Read the deltas, not the documents
Now you have something you could not have had by listening: a per-term view across time. Most terms will be flat. The few that moved are the article.
The term list that travels
Business-specific terms you will write yourself — product names, segment names, the metric this company gets asked about. These are the structural ones worth carrying into every set you build, because they describe how people talk rather than what they are talking about.
- Commitment verbs
- we will · we expect · we anticipate · we're targeting · we're comfortable with. These form a ladder from firm to soft. Movement down the ladder on the same subject across quarters is a specific, checkable observation.
- Time anchors
- by year end · in the second half · over time · in the coming quarters. A dated commitment becoming an undated one is a change in specificity even when every other word is identical.
- Hedges and qualifiers
- currently · at this point · as it stands · assuming · absent. Watch for these appearing next to a phrase that previously stood alone.
- Deflection patterns
- I'll let · we don't break that out · as we've said · too early to. In Q&A, who answers is data. A question that moved from one executive to another between quarters is worth noticing.
- Your own prior quarter's phrases
- Pull three or four distinctive stock phrases out of the oldest transcript in your set and search for them in the newest. A repeated phrase that suddenly stops appearing is invisible to every other method — you cannot notice an absence by listening.
What a delta looks like
Here is the shape of the thing you are looking for. This is an illustration of the pattern, not a real company's words.
Three things moved across four quarters and none of them is a number: a firm verb softened, a figure became a range and then disappeared, and a dated commitment became an undated one. What the sequence plausibly indicates is narrow and worth stating precisely — this is a company that has stopped putting a number and a date in the same sentence about margin. That is all it indicates. Reading any single quarter, all four sentences sound like confidence.
That is the entire argument for doing this in text. The Q4 sentence is not alarming on its own — it is a perfectly normal thing for an executive to say. It is only interesting next to Q1, and holding two sentences from twelve months apart in your head with enough precision to notice that "by year end" became "over time" is not a thing human memory does.
Keep prepared remarks and Q&A apart
Every call has two halves that deserve different reading. Prepared remarks are written in advance and reviewed before they are read aloud, so a change in them is deliberate — somebody chose the new wording, and somebody approved it. Q&A is what happens when the script runs out.
So mark the boundary in each transcript as you go — the operator's handover line makes it easy to find — and record which half each of your hits came from. A hedge appearing in prepared remarks and the same hedge appearing under a follow-up question are not the same observation, and a note that does not distinguish them is a note you cannot use in three months.
The Q&A half also carries something with no equivalent in any filing: which questions get asked repeatedly across quarters, and whether the answer changes. An analyst asking the same question for the third consecutive call is telling you the previous two answers did not land.
Queue four calls, come back to four transcripts
Paste a media URL or upload the file — timestamps and speaker labels included, exports as TXT, DOCX, SRT or JSON so the set drops into whatever you search it with. First 60 minutes free.
Start a transcript →Why the search has to be across the set
Searching inside one transcript is a convenience. Searching across four is a different capability, and it is the reason the set is the unit of work.
Export all four as plain text into one folder and every tool you already own becomes a research instrument: grep -n "by year end" *.txt in a terminal, or your editor's search-in-folder, or the search box in whatever notes app you live in. Line numbers and filenames come back with each hit, so you can walk straight to the quarter and the moment. No specialist software is involved, which is what makes the habit survive.
Keep the audio too. The transcript is the index; the audio is the evidence. When you find a sentence that matters, the timestamp takes you back to hearing it said — and tone, pause and hesitation are real information that no transcript captures. The text is how you find the thirty seconds worth listening to out of four hours.
Running this without spending anything
Worth saying plainly, because a workflow that only works if you buy something is a sales pitch with steps: every part of this method except the transcription itself is free, and a good deal of the raw material is free too.
- The filed text costs nothing. Every 8-K and its exhibits are on EDGAR, full-text searchable, permanently. For the quarters you cannot get audio for, the release is a real substitute for part of this exercise — you lose the Q&A, which is the good half, but the guidance language is there.
- The tools are things you already own. A text editor with search-in-folder, or grep in a terminal, is the entire toolchain. There is no software to buy for the comparison step, and the elaborate research platforms sold for this add very little over a folder of text files.
- Some companies post their own transcripts. Check the investor relations page before you transcribe anything. If the text already exists, use it — the same logic applies here as anywhere: never pay to recreate a document somebody already published.
- If you have a limited number of transcription minutes, spend them on the oldest call in your set, not the newest. The newest one is the most likely to be covered elsewhere in text; the four-quarters-ago call is the one nobody has written about since, and it is the anchor the whole comparison hangs on.
The honest constraint is transcription minutes, not the method. Everything else here is a habit and a folder.
What you should actually end up with
One page per set. Not a summary of each call — a summary of each call is a worse version of the transcript. The page holds: the four sources with links and dates, the term list you used, every term that moved with its four quotes and timestamps, and one line per moved term saying what you would need to check next.
That page takes a couple of hours to produce and stays useful for a year, because next quarter you do not start over. You add a fifth column, rerun the same list, and the work compounds. The version of this that does not compound — attend the call, take notes, move on — costs the same hour every quarter and leaves nothing behind.
Common questions
Are earnings call transcripts filed with the SEC?
Usually not. The earnings press release is commonly furnished as an exhibit to a Form 8-K under Item 2.02, Results of Operations and Financial Condition, and that is free on EDGAR. The call — the prepared remarks and the analyst Q&A — is a separate event, and most companies do not file a transcript of it. That gap between the written release and the spoken call is the reason the audio is worth working with at all.
Why are earnings calls publicly webcast in the first place?
Regulation FD requires that when an issuer discloses material nonpublic information to certain recipients, it also makes public disclosure — either by furnishing or filing a Form 8-K, or by a method "reasonably designed to provide broad, non-exclusionary distribution of the information to the public." The SEC's adopting release explicitly contemplated webcasting conference calls as such a method, and noted that a meeting open to the public but not otherwise webcast or broadcast electronically does not qualify. Open webcasts became the standard compliance path as a result.
How long do replays stay up?
It varies by company and no rule fixes it. Some investor relations archives go back years; others post a replay for a limited window or put it behind a registration form. Since you cannot tell in advance, capture on the day rather than assuming the file will wait for you.
What does a wording change actually tell you?
That a question exists — not what the answer is. A hedge appearing where there wasn't one might reflect genuine uncertainty, or a new speaker, or a rewritten script, or different counsel reviewing the remarks. The value is that you now have a specific, checkable question instead of a general impression. Answering it is separate work.
Do I need speaker labels?
On the Q&A half, yes. Who answers which question, and whether that changed between quarters, is an observation you can only make if the transcript distinguishes speakers. On prepared remarks it matters less, since the script is usually read in a fixed order.
Can I just use an AI summary instead?
A summary is the wrong output for this workflow, because the thing you are looking for is exactly what a summary removes. "Management expressed confidence in margins" is a faithful summary of all four quarters in the example above, and it deletes the only thing that was interesting. Use models on the full text if you like — ask one to list every hedging phrase, or to diff two quarters — but keep the verbatim transcript as the artefact you search.
Turn a season's worth of calls into a set you can search. Paste a link or upload a recording, get timestamped text back, and keep it. Credits never expire.
Transcribe a call →