Why Dividend Growth Figures Disagree Between Sites

Cognex has grown its dividend by 7.6% a year over five years. It has also shrunk it by 31.9% a year over the same five years.

Both figures come from the same payment record. Neither is an error. The difference is one decision about what counts as a dividend, and it is the largest of four decisions that sit behind almost every disagreement between two sites quoting the same company.

The four decisions

Every dividend growth figure is the output of four choices. Sites make them differently and rarely state them.

All figures as of 29 August 2026.

They are listed in order of damage. The first can reverse the sign of the answer. The last, which sounds like the most technical, turns out to barely matter, and we only know that because we measured it rather than assuming.

1. Whether a special dividend counts

This is the big one, and Cognex shows why.

Dividend Health Score

87Very safehigh confidence

Price

$62.97

Dividend yield

0.54%

Annual dividend

$0.34

Key statistics

Day range$62.07 – $63.14

52W range$34.60 – $72.88

Volume997.3K

Avg. volume2.1M

Dividend amount$0.34

P/E ratio, GAAP61.24×

Forward P/E31.11×

Beta1.50

Market cap$10.59B

Cognex has an 11-year payment record and 11 years of increases behind it. Its regular dividend went 0.225, 0.245, 0.265, 0.285, 0.305 and 0.325 USD across the last six complete fiscal years. That is a rise every single year, and it works out at 7.6% a year.

In the first of those years it also paid a special dividend of 2.00 USD, roughly nine times the regular payment. Count that as part of the 2020 dividend and the series starts at 2.225 USD and ends at 0.325 USD, which is a decline of 31.9% a year.

Dividend Health Score

78Safehigh confidence

Price

$459.96

Dividend yield

0.43%

Annual dividend

$2.00

Key statistics

Day range$449.09 – $470.00

52W range$197.00 – $805.75

Volume219.3K

Avg. volume312.0K

Dividend amount$2.00

P/E ratio, GAAP40.49×

Forward P/E38.04×

Beta0.61

Market cap$6.45B

Argan is the same shape. Its regular dividend went from 0.50 USD to 1.75 USD across five completed years, which is 28.5% a year. A 2.50 USD special in the first year turns that into a 10.2% annual decline.

Neither company cut anything. In both cases a one-off payment sits at the start of the window, inflating the base year, and every subsequent year is measured against a number that was never going to repeat.

Across our data, 31 companies have a special dividend landing on either end of the standard five-year window. For those, the two treatments differ by a median of 10.4 percentage points, and 28 of the 31 differ by more than 3. Sign reversals are common rather than exotic.

We exclude specials from growth and publish them separately. A special is not part of the payment a shareholder can plan around, and including one guarantees that the following year reads as a cut. Of the 798 instruments we cover, 325 have paid at least one, so this is not a rare edge case.

2. Which streak you are being shown

Two different numbers get called a dividend streak.

Years paid counts uninterrupted payment. Years increased counts consecutive raises. Of the payers publishing both, they disagree for the large majority, and where they differ the median gap is 17 years. A site quoting one number without saying which is quoting the flattering one more often than not.

This does not change a growth rate, but it changes the headline that sits beside it, and it is the single most common way two dividend pages describe the same company incompatibly. We cover the consequences in the aristocrats that are quietly shrinking, where 21 companies with a 25-year payment record are paying less than they did five years ago.

3. Which window

A growth rate is a comparison between two dates, and moving either end changes the answer.

We publish five, ten and twenty year figures for this reason. They routinely disagree with each other, and the disagreement is information rather than noise. 18 companies show a negative five-year rate alongside a positive twenty-year one. Another 31 show a positive five-year rate alongside a negative ten-year one, where a cut has aged out of the shorter window. The widest example reads +43.1% a year over five years and -16.6% over ten.

So two sites can both be right and differ by 60 percentage points, purely by choosing a different horizon. Any figure quoted without its window is uninterpretable, and a surprising number are.

4. Which year a payment belongs to, and why it matters less than you would think

This is the one we expected to matter and it largely does not.

Of the 798 instruments we cover, 138 do not close their financial year in December. Their payments therefore get sorted into years that do not line up with the calendar, most heavily at September, June and January year-ends. It looks like an obvious source of disagreement.

We tested it by rebuilding every non-December payer's dividend history on a calendar-year basis and recomputing the five-year rate both ways. The median difference is 0.13 percentage points. Only 13 of the 119 comparable companies differ by more than one point. For the overwhelming majority, fiscal-year bucketing is simply a relabelling of the same payments, and the growth rate is unchanged.

What does change is the label. For roughly half of the non-December payers, the fiscal year carrying a given payment is numbered one higher than the calendar year it fell in. So "FY2024 dividend" on one site and "2024 dividend" on another can describe two different twelve-month periods with entirely different amounts, even when both sites agree completely about the underlying payments.

The minority where the rate genuinely moves are companies whose fiscal boundary splits the payment stream unevenly, so one year holds three payments and another holds five. Carnival is the extreme case in our data, reading -14.6% a year on a fiscal basis and +14.9% on a calendar one.

When the payment record itself disagrees

The four decisions above all assume both sites are working from the same payments. Sometimes they are not, and that is the disagreement worth taking seriously.

Dividend data arrives from market data providers, and providers make mistakes: a payment recorded in the wrong currency, a special filed as a regular, a duplicated row, an amount that missed a share consolidation. None of these is exotic and none is visible from the outside. Two sites drawing on different providers, or on the same provider at different times, will simply hold different histories.

One group is worse than all the rest. Among the fifteen companies that pay in a currency they are not listed in, fourteen needed a correction, and converting their dividend into the listing currency flips the direction of one year-over-year comparison in nine.

We check every company's dividend record against the primary source rather than accepting what arrives. For US filers that means the company's own XBRL disclosures to the SEC; elsewhere it means the issuer's investor-relations history. Every one of the 798 instruments on the site has been through that check.

670 came back clean. 128 required a correction. That is 16% of the corpus where the record we received did not match what the company itself published, and where the figures you see are the corrected ones.

We mention this not to claim the others are wrong, because we cannot see their inputs. We mention it because it puts a floor under how much agreement is reasonable to expect. If roughly one company in six needs its dividend history corrected at source, then two sites showing identical figures for everything would be the surprising outcome, not the expected one.

It also changes what to do when you find a disagreement you cannot explain by window, streak, specials or year labelling. At that point the question is not which site did the arithmetic differently. It is which one matches the company's own filings, and that is checkable by anyone in about five minutes.

What we do

For the record, so that our own figures can be checked rather than trusted.

• Specials are excluded from total_dps and from every growth figure, and published separately as special_dps.
• Both streaks are published as separate fields, never merged into one number called "streak".
• Every horizon is published separately, and a horizon is left blank when the completed history does not reach back far enough, rather than filled from a shorter window.
• Fiscal years follow the company's own year end, which is why a label may not match a calendar year.
• The payment record is reconciled against the company's own filings, not accepted as delivered, and the corrections are re-applied on every sync rather than patched once.

There is one more thing worth stating, because it is the part that is easy to get wrong quietly. When a horizon cannot be honestly filled we leave it empty. A company with eight years of completed history gets a five-year growth rate and no twenty-year one, rather than the same eight-year figure printed three times under three different labels. That practice is invisible when it works and produces a blank cell, which looks like missing data and is the opposite of it.

None of those is the only defensible choice. The point is that they are stated, so a figure that disagrees with ours can be traced to a decision rather than to one of us being wrong.

How to compare two figures

• Ask what the window is before anything else. A five-year and a ten-year rate are not competing answers to one question.
• Check whether a special is in the base year. If a company paid a large one-off five years ago, expect the two sites to disagree violently, and expect the lower figure to be the misleading one.
• Establish which streak is meant. "25 years" answers two different questions and usually the more flattering one.
• Treat an unsourced figure as unusable. If a page does not say what it excludes, the number cannot be reconciled with anything, including itself a year later.

Our five-step check on whether a dividend will hold uses the payment history directly rather than a summary rate, which sidesteps most of this, and the screener exposes each horizon as its own field.