DEV Community: Heysupa

How to Use DeepSeek for SEO

Heysupa — Tue, 29 Sep 2026 06:47:41 +0000

The majority of articles regarding DeepSeek for SEO really discuss the launch of DeepSeek. This one is about the model's current source of income—live client employment.

Using DeepSeek as an SEO tool and getting DeepSeek to mention your company are two issues that are frequently confused. The first has a practical solution. I'll explain why the second hardly exists at all. One word of caution: all of the benchmarks below are from businesses that have products to sell, including DeepSeek.

What is DeepSeek, and is it a search engine?
No. DeepSeek is a chatbot that lacks an index and is based on open-weight language models. The web chat is free with an account and may be accessed via chat.deepseek.com, the mobile app, or an API. According to Momentic's Similarweb-based tracking, it had a 4.1% share of global AI-chatbot web visitors in May 2026, far less than ChatGPT's 53.9%.

What is an open-weight model? Open weights refers to the publication of the trained model file itself. It is available for download and can be used on any hardware. The majority of assistants keep the file secret and only grant you access via an API.

The whole point here is that separation. The 685-billion-parameter model of Hugging Face's current V3.2 release was released under the MIT license. There are two responses to the question "is DeepSeek free?" Both the chat app and the weights—which you can download and use for business purposes—are free.

In all honesty, we believe that the way search engines are framed in many articles is incorrect and leads consumers to search for a function that isn't present. When you pass text to DeepSeek, it reads it and analyses it. narrower work, truly competent at it.

What is DeepSeek actually good at for SEO?
Structured reasoning in large quantities for nearly nothing. According to DeepSeek's advertised price, the Flash model has a context window of one million tokens and costs $0.14 per million input tokens and $0.28 per million output. At that speed, you can cease rationing prompts, which alters what you bother to analyse, and push an entire Search Console export through it.

The jobs that it does well on for our clients are:

Grouping keywords. Paste 800 queries to obtain intent groupings, each of which has a rationale.

Investigation of the content gap. Give it both your and a competitor's sitemap URLs, and enquire about the topic families that only they cover.

Creation of a schema. When prompted to check, it detects its own syntactic mistakes and produces clean JSON-LD.

Look at the similarities between those. Every task has a checkable correct response, and a large context window is preferable to writing polish. A practical month-by-month DIY plan is a better place to start than any model if you're still figuring out which tasks are important and in what order.

Where does DeepSeek fall short?
Writing. Strong reasoning scores—70 on SWE-bench Verified and 82.4 on GPQA Diamond—are reported by DeepSeek's own model card; nevertheless, none of those benchmarks assess whether a paragraph reads like it was written by a human. In reality, its drafts are heavily headed, overly organised, and filled with the throat-clearing lines that an editor removes first.

Real-time data is the second gap. Since the normal API path lacks web access, any information it provides about a competitor's live pricing or a current SERP is recreated using training data. When you ask for a statistic, you'll frequently receive a confident figure with no supporting data.

Your data is the third gaps. No chat model can respond to the query "what should I fix first" because it is unable to view your Search Console or your current ranking. That's not a criticism of this one; it's true of all of them. The inputs that Google ranking lacks are covered in the beginner's guide.

Is DeepSeek safe to use for client work?
Unless you self-host it, do not use it for private content. The company will "directly collect, process, and store your Personal Data in People's Republic of China," according to DeepSeek's clear privacy policy, and inputs such as prompts and uploaded files are used to train its models. A way to opt out exists. By default, it's off.

A track record is also present. Wiz Research discovered an unauthenticated DeepSeek ClickHouse database in January 2025 that had more than a million lines of log data, including secret keys and chat histories.

Sort your input according to sensitivity. It's okay to share competition URLs, public keyword lists, and your own live blog outlines. Client contracts, unreleased pricing, and anything covered by an NDA: no. Run the MIT-licensed weights on your own infrastructure if you want the cost profile without the jurisdiction question. In exchange for someone maintaining it, the issue is resolved.

Can you get cited by DeepSeek, and does that matter?
No one can tell you the truth just yet. Every significant market citation analysis evaluates the AI capabilities of ChatGPT, Perplexity, Gemini, Copilot, or Google. Any tool that claims to have a DeepSeek visibility score is inferring it because none of them disclose DeepSeek citation data. In the meantime, the quantifiable share continues to concentrate: Profound discovered that between April 15 and June 30, Google's AI Mode citations increased by 8.4 times, making it the second most-cited domain in AI Mode.

Traffic is something you can quantify. Unless you create a unique channel group with one channel that matches deepseek.com on the session source and is positioned above the usual channels, DeepSeek sessions that arrive in GA4 under Direct or Referral by default disappear into noise.
When you report it, you should explicitly note two limitations: the count is low because desktop programs remove the referrer, and it is a trend line rather than a total.

In all honesty, DeepSeek is not where GEO work pays if you have limited hours, even though it costs us a feature to express this. According to Hall's engine-by-engine citation data, Copilot's top-50 source list extends to sites at DR 51, DR 38, and even DR 0, but the list for every other engine is DR 85 and higher. This makes Copilot the anomaly worth pursuing. There is really little chance for a mid-authority website to compete with ChatGPT's established players. What we track and what we don't are detailed on Klimb's DeepSeek integration page.

How should DeepSeek fit into a small team’s stack?
Never the front of it, but the cheap layer underneath. Use DeepSeek for bulk analysis, then write a better-written model and compare the outcome with your own Search Console results. Even the Pro tier ends up with a rounding error of almost an hour of your time, at $0.435 per million input tokens and $0.87 per million output.

Add numbers to it. The model expense is in single digits because clustering 2,000 queries and producing 40 briefs requires a few million tokens in total. The idea that runs across our DIY SEO guide for small businesses is that it costs you money to decide what to publish and then ship it.

One caution regarding client work. Thin AI drafts at volume are how websites lose ground, and cheap tokens entice you to create far more material than you can edit correctly. Set the cap based on editing capability rather than model price. DeepSeek for SEO won't help if your bottleneck isn't output but rather judgement.

On your own website, try it. Every nite, Klimb conducts a five-pillar audit of your website, converts each result into an assignable task with an easy-to-understand remedy, and verifies the lift using your own Search Console and GA4 instead of a third-party traffic estimate. It runs buyer-style questions against ChatGPT, Claude, Perplexity, and Google's AI Overviews to see whether you are named at all, and it recognises AI-referral sessions from DeepSeek, Gemini, Copilot, Grok, and Perplexity in GA4. You can disconnect at any moment, and access to your Google data is read-only. Take advantage of Klimb's 15-day trial, which is free and doesn't require a credit card.

Alcohol use disorder ICD-10 codes that get paid

Heysupa — Mon, 28 Sep 2026 06:27:25 +0000

This little thing takes a huge amount of time away from the pitch. After completing an evaluation and feeling comfortable with your formulation, you write "moderate alcohol use disorder, five criteria met." Exactly what your training required—precise, defendable.

There is nothing in your sentence to help someone in billing determine whether that is F10.10 or F10.20. They select one. If you ever find out, you learn which one they choose when a claim is returned.

What is the alcohol use disorder ICD-10 code?
The pattern you recorded will determine this. F10.10 represents uncomplicated alcohol abuse, F10.20 represents uncomplicated alcohol dependence, and F10.90 represents undefined alcohol use. Additional numbers are added by complications like withdrawal or drunkenness.

The weight is carried by the pattern digit, which is.1 for abuse,.2 for dependence, and.9 when no one spoke.

Why does DSM-5 severity not tell anyone which code to use?
due to the disparate logic that underpinned the two systems. According to the DSM-5, alcohol use disorder is a single condition that is rated according to how many of 11 criteria a person satisfies: mild is 2–3, moderate is 4–5, and severe is 6 or higher.

Abuse and dependence are two distinct patterns in ICD-10, neither of which has a criteria count associated with them. A documented rule that turns five conditions into a.2 does not exist.

What was altered in DSM-5? Using the same 11 criteria, it combined the previous abuse and dependence categories into a single alcohol use disorder, eliminating legal issues and adding yearning.

Thus, the gap is structural rather than the result of carelessness. You are billing in the older system and documenting in the newer one.

The four words that can solve it
Alongside your DSM-5 severity, name the ICD-10 pattern. It's the entire solution.

Insufficient: "Moderate alcohol use disorder, five criteria met."

More accurate: "Moderate alcohol use disorder, five criteria met; meets criteria for alcohol dependence."

The ambiguity is eliminated with four additional words. The code on the claim is the code you intended, no one calls you, and no one downstream guesses.

Instead, other organisations utilise a general approach that links mild to misuse and moderate and severe to dependence. That is still a local custom rather than a regulation from the rules, and it usually works. If your company employs one, put it in writing and follow it regularly since an auditor will want to know how you made the decision.

5 Common pitfalls that you can avoid
The assessment and the history don't agree. The customer misused alcohol in his twenties, according to your background. Your evaluation now indicates dependence. It's true that a coder might bill for both. Include the phrase "history of" in the past tense.

To be kind, code down. When a customer satisfies the criteria for dependence, F10.10 feels more compassionate and appears to a reviewer as a facility billing above its own diagnosis. Additionally, it's the quickest method for them to lose the residential authorisation you desired.

To secure the authorisation, code up. Because it shifts a payment issue into compliance territory, the opposite is worse. Write out your assessment.

After the evaluation, F10.90 will remain in place. Before you get the picture, unspecified is honest at intake. The once you have it, it ceases to be honest.

Forgetting that there are complicated digits. The record won't support the care you gave to a client in active withdrawal coded F10.20, who is described as calmer than they were.

When is F10.10 actually the right call?
The level of care you provide should be commensurate with the person's abuse criteria rather than their reliance. This code precisely explains outpatient treatment for a person whose drinking is producing serious issues without tolerance, withdrawal, or loss of control.

It is not inferior in any way. An accurate, lighter code is more defendable than an inflated, heavier one, and your customer does not benefit from the tendency to grasp for dependence because it seems more serious.

There is also a time when you are protected by the conservative code. If a person is admitted for evaluation but their reliance hasn't been determined yet, code what the chart supports now and update it after the issue has been resolved. Charts wind up contradicting their own promises when they code forward to what you expect to discover.

What a payer is actually reading in your notes
They want to know if someone who needs what you requested fits this condition. It is answered by F10.20 with a withdrawal issue. F10.90 provides no answers.

19% of in-network claims on marketplace plans were rejected by insurers in 2024; the percentage varied by insurer and ranged from 3% to 36%. Denial rates have been rising, and behavioural health is at the lower end.

Seldom will you recognise your own denial. When it gets to the person in charge of denials, they call you for records, and you have to spend forty minutes recreating a four-month-old session. A clause you write while the client is still in the room prevents that actual expense to you.

Where the diagnosis meets the authorization
Clinicians are surprised by one additional thing, so it's worth knowing. Some payers see it as an authorisation failure rather than a fixable coding issue if your organization received an authorisation citing one diagnosis and the claim is sent out with a different one.

This implies that a diagnostic you made in good faith between the authorisation call and the claim may result in a discrepancy that is not discovered until the remittance. It is not a justification for not improving the diagnosis. When you do, it's a good idea to let the person handling your prior authorisations know.

In 2024, less than 1% of rejected claims were appealed, and 66% of those appeals were upheld by insurers (KFF). The majority of what is challenged is lost, and very little is disputed.

How Supa handles the translation
Quiet errors are seen in handoff problems, which include the discrepancy between DSM-5 and ICD-10. Nobody is ever wrong. Simply said, the information is lost during the journey.

Supa operates within the systems you now utilise. It detects when an unidentified code appears on a chart intended for a residential authorisation, when the diagnosis on a claim differs from the one on the authorisation, and when your remark notes DSM-5 severity without mentioning an ICD-10 pattern.

You continue to use your clinical judgement. The version where your judgement was sound but the record didn't reflect it disappears.

How to Calculate Denial Rate: Methods Compared

Heysupa — Sun, 20 Sep 2026 12:14:04 +0000

The ops review was meant to be easy.

The billing vendor for your residential program reported a 6 percent denial rate for Q2. During the same period, your internal outpatient team reported 18%. Then, based on data from the EHR dashboard, your board packet displayed 11%. One organization, three "denial rates," three distinct narratives.

The CFO kept the rate constant in the budget for the following year. The revenue cycle director added two FTEs after noticing the increased figure. The board pressed to cut them when it noticed the lower vendor number. Since none of the three had the same denominator, numerator, or temporal basis, they were all "right."

This article's topics include:

Tiny worked examples of the primary methods used to compute the denial rate.

How the number is affected by claim-lifecycle (cohort) versus activity-based, claim versus line, count versus dollar, and initial versus total definitions.

Which definition should you use for staffing and daily operations, and which for your main KPI?

How to make your figures finally reconcile by standardising the denial rate across systems, vendors, and reports.

Why do denial rate reports conflict so often?
Fundamentally, the denial rate has to be straightforward: denials divided by a certain amount. In reality, practically every aspect of that statement is debatable:

What qualifies as a "denial."

whether you quantify lines or claims.

Whether you're counting dollars or events.

which date is the metric's anchor.

First-pass denials or any denials during the claim's duration are important to you.

Internal teams, billing providers, and various EHRs silently make different decisions. You only learn when:

Your EHR's "denial rate" appears to be half that of your vendor.

While your report utilised dollar-value, a benchmark you took from an HFMA webinar used claim-counts.
In this year's board deck, a "Q2 denial rate" was based on claim filing dates, whereas last year's was based on payment dates.

This is exacerbated for SUD and behavioural health by:

lengthy per diem claims with several lines.

Use various denial codes to separate payers.

frequent assessments of medical need and authorisation at the same time.

Denial rates will continue to lead to poor staffing, poor payer negotiations, and poor board discussions until you establish a single definition and direct all systems toward it. The glossary entry for denial rate on your intranet or in your BI tool exists because of a shared definition, which allows everyone to view the same information.

What are the main ways to calculate denial rate?
Denial rate definitions differ along five major axes:

Activity-based (remittance or incident) versus claim-lifecycle (cohort).

dollars-based versus count-based.

claim-level vs. line-level.

First-pass versus overall.

Date selection (submission, adjudication, and posting).
After focusing on the first four, we will discuss dates in the standardisation part.

What is claim-lifecycle (cohort) denial rate?
Definition: Track a group of claims filed over time. Ask yourself, "What share of this cohort was ever denied at least once at any point in its life?"

Tiny example:

100 claims were sent in July.

Over time, fifteen are denied eligibility, five are denied due to medical necessity, and three of those are eventually reversed on appeal.

July's cohort rejection rate is 20/100, or 20%.

Main points:

Even if a claim is rejected more than once, it can only be tallied once.

Until you are certain that remits have come and appeals have been settled, the figure is not final.

For root cause, you can divide by denial type, payer, program, or region.

What makes operators enjoy it:

It provides you with the actual "outcome" of your submission processes.

For end-to-end analysis, it works well with the clean claim rate and first-pass resolution rate.

Why it's difficult

It's slow. Particularly with Medicaid and appeals, denials on July submissions may take 60 to 120 days to complete.

Accurately building across several systems is more difficult.

What is activity-based (remittance / event) denial rate?
Definition: Regardless of when the claim was filed, examine every denial event that occurs during a time frame. Compare these with the lines or claims handled during that time.

A small example

Your team posts 400 claims totalling remits for all payers and service dates in July.

Fifty denial incidents are included in those remits (some are repetitions of the same assertion).

July's activity-based denial event rate is 50/400, or 12.5%.

This is what the majority of clearinghouse and EHR "denial dashboards" default to. Instead of looking at the original submission dates, they examine 835s that were uploaded or adjudication dates within a period.

What makes operators enjoy it:

Almost real-time. This week, a payer system disruption is visible.

This month's workqueue is experiencing denials, which indicates a staffing pressure.

Why it is distorted

It counts twice. A claim that is rejected, amended, and then rejected once more may be counted twice.

Cohorts are mixed. Submissions from April, May, and June may be the source of July's activity.

More difficult to compare using cohort-based benchmarks.

What is count-based vs dollar-based denial rate?
Count-based: counts of claims or lines make up the denominator and numerator. Dollar-based: Dollar quantities (charges or anticipated reimbursement) make up the denominator and numerator.

A small example

Rejected in July
Ten denials of claims. $10,000 is the total amount of refused charges.
The total amount of charges you filed for the time was $200,000.

Claim denial rate depending on counts:

10% is the ratio of 10 rejected claims to 100 submitted claims (sample numbers).

Denial rate in dollars:

10,000 dollars rejected out of a total of 200,000 dollars submitted is 5%.

Why the distinction matters

While paying the premium per-diem without any problems, a payer may reject several small group therapy codes. This results in a considerably lower dollar-based denial rate and a high count-based denial rate.

For medical necessity, a different payer may consistently reject expensive residential days. As a result, while counts remain stable, the dollar-based denial rate rises.
In reality, boards and CFOs are concerned with dollar-based. Front-line RCM teams perceive their workload to be count-based.

What is line-level vs claim-level denial rate?
Claim-level: The entire claim is deemed refused if any line on it is rejected. Line-level: the percentage of all service lines that were rejected is calculated by counting each line individually.

Small example (one assertion):

Five lines in one claim:

Four paid per-diem days.

Due to "no auth on file," one per diem day was rejected.

At claim level:

For metric purposes, this claim is "denied."

This small sample's claim-level count-based denial rate is 1.

100% is the ratio of denied claim to total claim.

Line-wise:

Twenty percent is one refused line out of five.

This distinction is crucial to behavioural health:

IOP and PHP claims sometimes combine several days or services.
Because authorisation spans for residential per-diem frequently vary, some days experience auth problems while others do not.
Claim-level is acceptable if all that matters is whether any part of the claim has to be reworked. Line-level data is more precise when it comes to payer behaviour or use.

What is initial (first-pass) vs total denial rate?
Initial (or first-pass) denial rate: was this claim or line initially denied by the payer?

Total denial rate: Was there ever a refusal of this claim or line, even after adjustments or appeals?

A small example

100 claims were filed in June.

Initial answers:

15 are rejected.

85 pay or transition to a paid/zero-pay position without being rejected.

Of the 15 rejected:
Ten are fixed, and the second submission is paid.
Five are rejected once more before being written off.

Initial rate of denial:

Fifteen percent (15 initial denials out of 100 submissions).
Total denial rate (depending on cohort):

Out of 100 submissions, 15 claims received at least one denial, or 15%.

If "total" were defined as "ever denied, including repeats," you could count 20 denials for each of those 15 statements. Double counting appears at that point.

First-pass denial rate and first-pass resolution rate are closely related. It provides information about front-end quality, including eligibility, demographics, authorisations, and codes. The total denial rate, which correlates with FTEs and burnout, indicates how much work affected your denial work queues.

How do different setups change the “right” denial rate definition?
Program type, payer mix, and billing methodology all play a major role in your optimum definition.

Clinics for outpatients with brief episodes

Features that

One or a few CPT codes for each visit.

brief incidents, frequently just one DOS per allegation.

Low average charge per claim due to high volume.
Repercussions

Because most claims contain one or two lines, claim-level and line-level are frequently comparable.

The count-based denial rate keeps a careful eye on how frequently your team handles a claim.

What's effective:

Use the initial, dollar-based, claim-level denial rate as your primary KPI. It is simple to convey to payers and leadership.
activity-based count of staffing denial occurrences, as denials are received and handled promptly.
Take caution:

Denials based on eligibility may increase by employer group or state program. For true insight, break down the denial rate by payer or product.
Carve-out plans for behavioural health may generate peculiar rejection codes that are difficult for your EHR to map. These must be normalised by your denial-code library.

Residential, PHP, and IOP with long per-diem episodes
Characteristics:

Claims are sometimes rolled across lengthy episodes, with multiple days per claim.

excessive dependence on concurrent and initial authorisation.
Some days are paid, while others are denied for "no auth," "not medically necessary," or "level of care not covered." These partial denials are frequent.

Repercussions

The story is drastically altered by line-level versus claim-level. A claim can be flagged for a single refused day.
Because rejected days at high ASAM levels can occasionally ruin profitability, dollar-based metrics are more important.
What's effective:

Initial, claim-level, dollar-based cohort denial rate by payer and care level should be your primary KPI. It clearly illustrates financial danger.

Line-level denial analysis by denial cause, including utilisation and medical necessity, is important for clinical and UR teams.

Because each refused day frequently results in interdisciplinary labour across UR, clinical, and billing, staffing should employ an activity-based count of denied lines.

Take caution:

An auth problem for days 8 through 10 appears to be the same as a complete denial for all days if you only report the claim-level denial rate. You lose important nuance when discussing auth policies.
The way that state Medicaid MCOs encode authorisation and medical-necessity denials varies greatly. That has to be handled state-by-state in your mapping.

Single-state vs multi-state behavioral health
Single-state:

Normalising denial codes and categories is made easier.
The behaviour of the Medicaid program is consistent.

Multi-state:

Medicaid and MCO editing regulations vary greatly.
Different combinations of CARC and RARC may be used for the same real-world denial in each state.

Significance:

For the metric, you want a single master definition of "denial," but your denial-code library should include adaptable payer-specific mappings.
You can see when a state's program modifies its regulations in the middle of the year by using activity-based indicators.

Conclusion:

Ensure that the KPI definition is consistent across states. The technique, date basis, numerator, and denominator are the same.
Dashboards should be divided by payer and state to analyse variations.

In your metric specification, note exceptions that are unique to a state.

In-house billing vs outsourced billing vendors
Internal groups:

If you take the time to standardise, you will have greater control over definitions.

Risk: Each of IT, finance, and billing may develop a denial metric of their own.

Outsourcing suppliers:

Most bring their own definition of denial rate, which is frequently activity-based and claim-count orientated.

Your finance system or EHR may not work with their gateway.

How to do it:

Make your standard definition public and mandate that vendors map to or comply with it.

When requesting contracts and RFPs, ask vendors for:

precisely which denominator and numerator they employ.

They may count funds, count lines, or make claims.

Ignoring this will put you back in the "6 percent vs. 18 percent" predicament, with no clear means to reconcile and no ability to see how well your vendors are performing.

Which denial rate should you manage to?
Five denial rates are not necessary for each board packet. One main "managed number," along with a few operational partners, are required.

For the majority of SUD and behavioural health organisations, I advise:

Primary KPI: Initial, claim-level, dollar-based, cohort denial rate by submission month.

Numerator: Total denied permitted dollars for claims filed throughout the period that receive a rejection as their initial adjudication answer.

Denominator: The total anticipated amount of money permitted on all claims filed at that time.

Cut by: Program, payer, care level, and occasionally geography.

Why

Dollar-based is closely linked to income.
Front-end quality (eligibility, auth, coding) is reflected in initial denials, which are genuinely fixable.
Budgets and quantities are scheduled on a cohort-by-submission-month basis.

Operational companion: Activity-based denial event rate (count-based, claim-level) by posting month.

Denial events posted during the period are the numerator.
Denominator: The total number of claims during the time with remits posted.

Why

This aids with staff denial resolution and informs you of the number of things that enter the work queue.
It is almost instantaneous.

Supplemental: Line-level denial analysis for residential, PHP, and IOP.

Counts and dollars by refusal reason, per-day, or per-service may be displayed instead of a single "rate."
drives discussions about authorisation and documentation in clinical and UR settings.

In every report, you should be clear:
"This is the initial cohort denial rate by submission month, dollar-based, and claim-level."This is based on activity, event count, denial rate at the claim level, and remit-posting month.

The fact that the two numbers are different becomes normal and not stressful when they are properly labelled.

How do you standardize denial rate across your organization?

Three components make up standardisation: one owner, one temporal basis, and one definition.

How do you nail the numerator and denominator?
Put this in a plain-language specification that your RCM and finance departments can use.

As an example:

Metric designation: "Initial denial rate (dollars, claim-level, cohort)."

Dividend:
any claim made throughout the time frame.
Don't include assertions that are only informational or free.
If you have that reasoning in place, use the estimated authorised amount (not full charges) as the dollar base.

Calculator:

subset of denominator in which at least one rejection code (CARC with denial status on the 835) that lowers payment is present in the initial adjudication response.
As long as at least one line is refused, include claims with partial payments.

Don't include "soft edits" that don't involve staff work, such as informational messages that don't affect payments.

Next, describe "denial" in terms of codes:

Consult the CARC and RARC codes associated with your centralised denial-code library that you believe to be true denials.

Choose whether to incorporate:

Rejections of medical necessity and use.

Denials of eligibility and benefits coordination.

Services not covered.

Denials of repeated claims.

Be specific. "We include CARC codes in categories X, Y, Z. We exclude duplicates and informational edits."

How do you fix the date basis?
The date fields that power your metric must be selected:

Use the 837 or your billing system's submission date for cohort metrics.
Use the adjudication date if you publish in real time or the remit-post date in your PM/EHR for activity metrics.
Reasons to care:

A "Q2 denial rate" may refer to distinct claim populations if one report truncates according to the submission date and another according to the payment date.
Processing delays for state Medicaid programs may be lengthy. Your time series won't be distorted by those delays thanks to a submission-based cohort.
Put this in your specification: "For this metric, claims are attributed to the month they were first submitted to the payer."

How do you document shared definitions?
Make a basic internal "metrics catalogue" with the following:

A brief description and the metric name.
Details about the denominator and numerator.
Count versus dollar-based; claim against line-level.
First versus overall.
time-based.
sources of data (data warehouse, clearinghouse, Supabill, billing platform, EHR).
Connect this to your lexicon:

Denial rate link.
Since the first-pass resolution rate and clean claim rate are located in the same area of the funnel, cross-reference them.
Instead of using their preferred BI tool, anyone gathering data for the board or a payer should begin there.

Who owns denial rate and what cadence works?

Identify a clear owner, typically the VP of Revenue Cycle or the RCM director.

Establish a reconciliation rhythm:

Owner evaluates the denial rate every month across:
PM or EHR system.
Warehouse/BI.
ports for vendors.
applies the standard definition to compare them.
records and addresses the reasons for variance:
date filters.
inclusion or exclusion of particular programs or payers.
Disparities between claims and lines.

Quarterly

Give the board and executive team a one-page "denial rate definition and sources" PowerPoint.
Verify the definitions of any external benchmarks that are utilised, or clearly identify any discrepancies.

How do you force vendors to align or map?
Although you might not be able to compel every external party to alter their internal reports, you have two choices at your disposal:

Preferred: The vendor creates a report that precisely fits your requirements.

Acceptable backup: You calculate the measure in your own environment using the raw data that the vendor provides (claim-level with denial flags, dollars, dates, and CARC/RARC codes).

Add this to scorecards and contracts:
"The vendor shall either supply comprehensive claim-level extracts that enable the client to calculate it or monthly denial reports that complies with the client's metric demand."
Demand a monthly side-by-side table that maps the vendor's measure to yours if they insist on quoting "their" denial rate.

How can AI help keep denial rate definitions consistent?

An agent may reconcile claim-by-claim, tag mismatches, and generate a single source-of-truth denial rate if your EHR, clearinghouse, and data warehouse don't agree.

Whether billing is done internally or externally, Supabill uses a single, uniform metric definition for all payers and programs, then presents it to clients in the same manner. In order to ensure that your internal and vendor numbers match, Supabill's agents monitor remits, tag rejections with your standardised reasons from the denial-code library, and compute both cohort and activity-based denial metrics.

One truthful limit:

Bad or missing source data cannot be fixed by AI. An AI agent still requires human judgement to create the appropriate business rules or to initiate manual review if authorisations are not recorded, if employees select arbitrary refusal reason codes, or if payers submit incomplete 835s. Agents do the counting, while humans maintain the definition and governance of metrics.

Billing for Treatment Centers: In-House, Outsourced, or Automated?

Heysupa — Sun, 20 Sep 2026 09:47:40 +0000

The query "should we keep this in-house or outsource it?" is older than the majority of the codes you bill if you manage billing for a behavioural health treatment facility. The opposite side of the table has changed.

Your claims are no longer decided by hand by payers. By appealing one claim at a time, they put them through automated algorithms that may reject them in a matter of seconds at a scale that no billing team can match. This change does more than just make denials more obnoxious. It modifies the revenue cycle's staffing and tooling calculations. When compared to a slow, human payer, a billing mechanism that was "good enough" can quietly lose six figures annually to an automated one.

The operator making that decision—an owner, COO, or RCM lead at a multi-site IOP, PHP, residential, or SUD program with a messy payer mix and actual claim volume—is the target audience for this guidance. It's not for the lone therapist who is debating whether or not to study modifier rules. We'll outline the three actual models, their true costs, the situations in which each actually prevails, and how the automated-payer transition should affect the outcome. Not at all. Just the trade-offs in their true form.

What billing actually costs a treatment center at volume

Prior to comparing models, be truthful about your current expenses. The majority of the expense is not visible on the billing line item.

Denials that are never redone, days in A/R that tie up cash, soft denials (silent downcoding) that no one tracks, write-offs that are filed after the deadline, and staff time spent battling all of it are the five areas where a treatment center's revenue leaks simultaneously. All of these have a strong connection to behavioural health. Compared to a medical baseline of 90–95%, practices receive 68–85% clean claims. That 15–30% discrepancy is revenue that is constantly in the air.

At scale, the math complicates quickly. For every 1% decrease in the denial rate, each full-time clinician makes between $4,000 and $8,000 annually. A rounding error is not being recovered by a 20-clinician center that moves from 18% to 10% denials. It's a hire recovery. Due to their high prior-auth and concurrent-review requirements, the level-of-care services that characterise treatment centers (IOP, PHP, and residential) are denied at the highest rates in behavioural health, 20–30%. Your billing model is therefore not a back-office option. It's one of your margin's bigger levers.

The payer side is industrializing denials

This is the element that is actually new and explains why a billing structure from the 2019 timeframe is riskier than it appears.

In recent years, payers have automated the adjudication process. Take the most well-known example. According to an inquiry, Cigna rejected over 300,000 claims in two months using an automated system called "PxDx," which took an average of 1.2 seconds per claim. Physicians signed off on batches without opening patient files. According to one former reviewer, "we just click and submit." 50 at a time can be completed in about ten seconds.

There are multiple payers. A class-action lawsuit was filed against UnitedHealth for using an algorithm called nH Predict to refuse post-acute care. The lawsuit claimed a 90% error rate, which means that when patients resisted, about 90% of its denials were reversed. Major Medicare Advantage insurers relied on automation to reject prior authorisation for post-acute care, according to a 2024 U.S. Senate Permanent Subcommittee on Investigations report.

The strain on the supplier side is quantifiable. According to the American Medical Association's annual pre-authorization poll of practicing physicians, 95% of respondents indicate that prior authorisation delays care, 79% say that it causes patients to stop receiving treatment, and 26%—more than one in four—say that a prior authorisation delay resulted in a significant adverse event. It's an incredible volume. In 2024, Medicare Advantage insurers alone made about 53 million prior-authorization decisions. The industry as a whole is seeing an increase in denial rates rather than a decrease.

Automation isn't always the bad guy, to be honest. Not all denials are incorrect, and the same techniques can expedite the approval of valid claims. The imbalance is the issue. While most providers still file appeals by hand, one fax at a time, payers make decisions at machine size. Bringing equal leverage to your side of the cycle is the only long-term solution. This is the perspective to apply to the following three models: Which one enables you to combat an automated payer without exhausting your billing staff?

The three models for treatment-center billing

After removing the marketing, a treatment center's revenue cycle can be managed in three different ways:

Internal RCM team: you recruit, educate, and provide tools for billers, programmers, and AR experts.

Outsourced RCM/billing company: billing for a portion of collections is handled by a third party.

Agentic billing platform: software that uses agents trained in behavioural health to manage the rules-based tasks, while your staff oversees the edges.

There is no conflict between these. In-house front-end eligibility, outsourced back-end appeals, or a platform supporting a lean internal team are examples of how many centers combine these. However, the dominant model determines your rejection performance ceiling and cost structure. To be honest, let's handle each one separately.

Model 1: In-house RCM team

When it wins: it will have a high, steady number of claims, a concentrated payer mix that you are familiar with, leadership that demands complete control and visibility, and sufficient scale to cover redundancy in the event of a resignation.

The closest feedback loop is provided by an internal staff. A reoccurring documentation gap can be resolved in a team meeting rather than a support ticket because your billers are seated down the hall from your clinicians. The institutional memory of which Medicaid MCO downgrades which code, as well as the data and relationships with your major payers, belong to you.

The salaries are lower than the actual loaded cost. One biller is not a treatment-center RCM function. Eligibility/auth, charge entry, coding, follow-up on claims, denials/appeals, and patient billing are all handled by three to eight individuals, depending on volume. Increase managerial costs, training to stay current with payer changes, and software (clearinghouse, scrubbing, reporting). Add the expense that no one plans for: turnover. Roles in billing change. Your appeal backlog and A/R days increase over the two months it takes to rehire and ramp up after your denials specialist departs. That gap is costly for an auth-heavy IOP/PHP center.

On its own, In-house also faces challenges with the new payer realities. The speed at which an automated payer rejects a claim is simply too fast for a human team to scrub every claim against every payer's current guidelines across 50 states.

Model 2: Outsourced RCM / billing company

When it wins: You want to turn a fixed cost into a variable one; your internal function is failing and you need a capable operator immediately; or you're growing swiftly and can't hire RCM talent fast enough.

A competent billing company starts out with a skilled staff, pre-existing payer contacts, and procedures. Usually, you pay between 4 and 9% of collections (often more for complex behavioural health), which is not fixed payroll but rather scales with income. That may be a quick improvement for a clinic that is losing money due to an overworked internal staff.

The drawbacks are genuine and worth mentioning:

Incentive mismatch. A company that bills based on collections maximises its profit margin. It may not pursue the hard, low-dollar denials that are costly to appeal—exactly the ones that stack up at a treatment facility—and will gladly collect the easy 80%.

Diminished visibility. Your denial information is stored in their system. Your clinicians do not receive the close feedback loop that stops rejections upstream, and you discover a systemic issue later.

Behavioral-health particularity. The level-of-care, concurrent-review, and parity details that influence your denials are often overlooked by typical medical billing businesses, which consider BH as an afterthought.

One solution to a staffing issue is outsourcing. It does not resolve the automated-payer issue on its own. You are paying a percentage to hire people to battle robots on your behalf.

Model 3: Agentic billing platform

When it wins: You have a complicated, multi-state payer mix with continuously changing regulations; you want to maintain control and visibility while matching the automation of the payers; or you want to significantly increase the productivity of a lean in-house staff rather than increasing manpower linearly with volume.

The high-volume, rules-based tasks—eligibility checks, benefit and auth verification, claim scrubbing against payer-and-state regulations, rejection routing, and appeal drafting—are continually carried out by an agentic platform. Over time, it learns your unique payment patterns. Up the value chain, your team encounters oversight, exceptions, and the decisions that software shouldn't make.

Any vendor who refuses to disclose their honest limits is trying to sell you something. A payer's policy cannot be changed by a platform. Good software detects the change and stops submitting when a payer changes a code to non-covered, but it cannot force the payer to pay. It cannot correct an incorrect diagnosis or bill for a clinically insufficient session. Furthermore, it won't fix a flawed clinical workflow. The cross-team, cross-payer, and cross-state work that people make mistakes on because there is too much of it to store in working memory is what it corrects. The automatic payers are taking use of exactly that effort.

The hidden costs nobody quotes you

It is incorrect to compare a biller's income with the "7% of collections" that a billing company claims. The revenue that a billing model is unable to generate is its true cost:

Even if the claim is subsequently paid, denial rework costs between $25 and $70 each claim in staff time.

Days in A/R: money you have earned but are unable to utilise. This decreases by five to ten days with same-day note completion and clean first-pass claims; it increases with a slow model.

A 90837 was paid at the 90834 rate due to soft denials or silent downcoding. Most centers never monitor the $2,400–$5,000 lost per 1,000 claims when 8–10% of claims are subtly decreased by $30–$50.

Clinicians seeing patients before they are in-network with your payers is known as credentialing lag, and it results in sessions that are not billable.

Timely-filing write-offs: claims that expire after 90–180 days and are not appealed.

The cost of a denial that advances to the appeal stage is 5–10× higher than that of a denial that was stopped prior to filing. Rather than the model with the lowest sticker price, the model that prevents the most denials upstream wins.

A decision framework

It comes down to three questions. What is your volume of claims? (Rarely does low volume warrant a whole internal staff.) To what extent is your payer mix flexible and complex? (More intricacy encourages automation over manual procedures.) To what extent is your internal RCM expertise developed? (Outsourcing or a platform is safer when strong leadership is absent; strong leadership makes in-house plus tooling stronger.)

AI and agentic systems in behavioral health billing

When you take a step back, it becomes evident that denials are almost never a single point of failure. A refusal can be traced upstream to a payer rule that changed last quarter, a code that didn't match recorded time, a benefits check that missed a session restriction, or a note that didn't prove medical necessity. In order to fix denials, all of those locations must be fixed simultaneously, continuously, and at the current speed at which payers operate.

That is the concept behind Supabill. It isn't just one tool for billing. Throughout the whole revenue cycle, an integrated system of agents with training in behavioural health collaborates around the clock. Before the initial visit, the benefits verification agent extracts eligibility and reveals session restrictions, prior-auth requirements, and cost structure.

It interacts with more than 5,000 payers, and if a portal doesn't respond, it can really call the payer to obtain the benefits. The claims-scrubbing agent flags the patient's submissions as they get closer to visit 18, before an avoidable denial, when it discovers that the payer caps psychotherapy at 20 sessions annually. State-by-state and payer-by-payer regulations are stored in the claims agent's core database, and each claim is pre-submitted for time-vs-code, modifier, POS, prior-auth match, and bundled-code discrepancies. The denials-management agent receives anything that is still rejected, routes it by payer and code, writes appealing against relevant clinical contexts, and keeps track of resubmission cycles.

Additionally, when that agent discovers that a payer is rejecting a code for a certain missing element, that information feeds back into Supanote's documentation guardrails, causing the note to prompt for it the next time. Every agent makes the others more intelligent. Once a denial reason is learned, it becomes universally preventable.

It's not about replacing your crew. In order to prevent treatment centers from bringing a stopwatch to an algorithmic battle, it is intended to provide the provider side the same leverage that the payer side already possesses. Instead of performing the repetitive, cross-payer rule-checking that individuals frequently make mistakes on, your staff should focus on the judgement job that truly requires them.

Again, the honest restrictions are that agents cannot correct an incorrect diagnosis, alter payer policy, or restore a malfunctioning clinical workflow. They address the cross-team, high-volume, rules-based job that is precisely where automated payers are already succeeding.

Quick wins you can act on this quarter

Determine the actual cost of rejection. Get the denial rate, average denial amount, days in A/R, and an estimate of soft-denial leakage by comparing the actual and predicted reimbursement per CPT for the previous quarter. The figure is the starting point for any model decision and is typically higher than leadership believes.Examine the reasons behind your previous 50 denials (prior authorisation, coding, eligibility, medical necessity, paperwork, timely submission, payer policy). The shape indicates if you have a personnel issue, a tooling issue, or both.Examine any platform or vendor's behavioural health level of care under pressure. Enquire especially about how they manage ASAM medical necessity and IOP/PHP concurrent review. Generic responses raise suspicions.

Verify eligibility and authentication again 24 to 48 hours prior to each visit. Regardless of the model you use, it is the one process modification that closes the most front-end denials.Monitor not only the denial rate but also the first-pass clean claims rate as a key KPI. HFMA's goal is 95%+; behavioural health centers reach 92–95% with structured documents and pre-submission cleaning.